Whalswp commited on
Commit
e68e388
·
verified ·
1 Parent(s): 21c514c

Add files using upload-large-folder tool

Browse files
Files changed (41) hide show
  1. .gitattributes +3 -0
  2. RKDv3/base120_p2-40k_p3-80k_fm1-rkd0p5_atomic65_human.stopped_nospace_20260708_204923/default/base120_p2-40k_p3-80k_fm1-rkd0p5_atomic65_human.yaml +49 -0
  3. RKDv3/base120_p2-40k_p3-80k_fm1-rkd0p5_atomic65_human.stopped_nospace_20260708_204923/default/phase2/experiment_cfg/metadata.json +431 -0
  4. RKDv3/base120_p2-40k_p3-80k_fm1-rkd0p5_atomic65_human.stopped_nospace_20260708_204923/default/resolved_config.yaml +50 -0
  5. RKDv3/base120_p2-40k_p3-80k_fm1-rkd0p5_atomic65_human/default/base120_p2-40k_p3-80k_fm1-rkd0p5_atomic65_human.yaml +49 -0
  6. RKDv3/base120_p2-40k_p3-80k_fm1-rkd0p5_atomic65_human/default/phase2/config.json +101 -0
  7. RKDv3/base120_p2-40k_p3-80k_fm1-rkd0p5_atomic65_human/default/phase2/experiment_cfg/metadata.json +431 -0
  8. RKDv3/base120_p2-40k_p3-80k_fm1-rkd0p5_atomic65_human/default/phase2/model-00001-of-00002.safetensors +3 -0
  9. RKDv3/base120_p2-40k_p3-80k_fm1-rkd0p5_atomic65_human/default/phase2/model-00002-of-00002.safetensors +3 -0
  10. RKDv3/base120_p2-40k_p3-80k_fm1-rkd0p5_atomic65_human/default/phase2/model.safetensors.index.json +0 -0
  11. RKDv3/base120_p2-40k_p3-80k_fm1-rkd0p5_atomic65_human/default/phase2/trainer_state.json +0 -0
  12. RKDv3/base120_p2-40k_p3-80k_fm1-rkd0p5_atomic65_human/default/phase2/training_args.bin +3 -0
  13. RKDv3/base120_p2-40k_p3-80k_fm1-rkd0p5_atomic65_human/default/phase3/config.json +101 -0
  14. RKDv3/base120_p2-40k_p3-80k_fm1-rkd0p5_atomic65_human/default/phase3/experiment_cfg/metadata.json +431 -0
  15. RKDv3/base120_p2-40k_p3-80k_fm1-rkd0p5_atomic65_human/default/phase3/model-00001-of-00002.safetensors +3 -0
  16. RKDv3/base120_p2-40k_p3-80k_fm1-rkd0p5_atomic65_human/default/phase3/model-00002-of-00002.safetensors +3 -0
  17. RKDv3/base120_p2-40k_p3-80k_fm1-rkd0p5_atomic65_human/default/phase3/model.safetensors.index.json +0 -0
  18. RKDv3/base120_p2-40k_p3-80k_fm1-rkd0p5_atomic65_human/default/phase3/trainer_state.json +3 -0
  19. RKDv3/base120_p2-40k_p3-80k_fm1-rkd0p5_atomic65_human/default/phase3/training_args.bin +3 -0
  20. RKDv3/base120_p2-40k_p3-80k_fm1-rkd0p5_atomic65_human/default/resolved_config.yaml +50 -0
  21. RKDv3/base120_p2-60k_p3-60k_fm1-rkd0p5_atomic65_human.stopped_nospace_20260708_204923/default/base120_p2-60k_p3-60k_fm1-rkd0p5_atomic65_human.yaml +49 -0
  22. RKDv3/base120_p2-60k_p3-60k_fm1-rkd0p5_atomic65_human.stopped_nospace_20260708_204923/default/phase2/experiment_cfg/metadata.json +431 -0
  23. RKDv3/base120_p2-60k_p3-60k_fm1-rkd0p5_atomic65_human.stopped_nospace_20260708_204923/default/resolved_config.yaml +50 -0
  24. RKDv3/base120_p2-60k_p3-60k_fm1-rkd0p5_atomic65_human/default/base120_p2-60k_p3-60k_fm1-rkd0p5_atomic65_human.yaml +49 -0
  25. RKDv3/base120_p2-60k_p3-60k_fm1-rkd0p5_atomic65_human/default/phase2/config.json +101 -0
  26. RKDv3/base120_p2-60k_p3-60k_fm1-rkd0p5_atomic65_human/default/phase2/experiment_cfg/metadata.json +431 -0
  27. RKDv3/base120_p2-60k_p3-60k_fm1-rkd0p5_atomic65_human/default/phase2/model-00001-of-00002.safetensors +3 -0
  28. RKDv3/base120_p2-60k_p3-60k_fm1-rkd0p5_atomic65_human/default/phase2/model-00002-of-00002.safetensors +3 -0
  29. RKDv3/base120_p2-60k_p3-60k_fm1-rkd0p5_atomic65_human/default/phase2/model.safetensors.index.json +0 -0
  30. RKDv3/base120_p2-60k_p3-60k_fm1-rkd0p5_atomic65_human/default/phase2/trainer_state.json +3 -0
  31. RKDv3/base120_p2-60k_p3-60k_fm1-rkd0p5_atomic65_human/default/phase2/training_args.bin +3 -0
  32. RKDv3/base120_p2-60k_p3-60k_fm1-rkd0p5_atomic65_human/default/phase3/config.json +101 -0
  33. RKDv3/base120_p2-60k_p3-60k_fm1-rkd0p5_atomic65_human/default/phase3/experiment_cfg/metadata.json +431 -0
  34. RKDv3/base120_p2-60k_p3-60k_fm1-rkd0p5_atomic65_human/default/phase3/model-00001-of-00002.safetensors +3 -0
  35. RKDv3/base120_p2-60k_p3-60k_fm1-rkd0p5_atomic65_human/default/phase3/model-00002-of-00002.safetensors +3 -0
  36. RKDv3/base120_p2-60k_p3-60k_fm1-rkd0p5_atomic65_human/default/phase3/model.safetensors.index.json +0 -0
  37. RKDv3/base120_p2-60k_p3-60k_fm1-rkd0p5_atomic65_human/default/phase3/trainer_state.json +3 -0
  38. RKDv3/base120_p2-60k_p3-60k_fm1-rkd0p5_atomic65_human/default/phase3/training_args.bin +3 -0
  39. RKDv3/base120_p2-60k_p3-60k_fm1-rkd0p5_atomic65_human/default/resolved_config.yaml +50 -0
  40. RKDv3/base120_single_fm1-rkd1to0_cosine_atomic65_human/default/base120_single_fm1-rkd1to0_cosine_atomic65_human.yaml +3 -3
  41. RKDv3/base120_single_fm1-rkd1to0_cosine_atomic65_human/default/resolved_config.yaml +3 -3
.gitattributes CHANGED
@@ -58,3 +58,6 @@ TimewarpVAE/logs/RKDv3_TimewarpVAE_DA_APHead_FM_RKD_CosineDecay_from_Atomic26Bas
58
  TimewarpVAE/RKDv3_DA_APHead_FM_RKD_CosineDecay_from_GR00TBase_to_120k/default/fm_rkdv3_timewarp_vae_da_aphead_cosine_decay_120k/checkpoint-90000/trainer_state.json filter=lfs diff=lfs merge=lfs -text
59
  TimewarpVAE/RKDv3_DA_APHead_FM_RKD_CosineDecay_from_GR00TBase_to_120k/default/fm_rkdv3_timewarp_vae_da_aphead_cosine_decay_120k/trainer_state.json filter=lfs diff=lfs merge=lfs -text
60
  TimewarpVAE/logs/RKDv3_TimewarpVAE_DA_APHead_FM_RKD_CosineDecay_from_GR00TBase_to_120k_gpu7_bs128_numgpu1_wandb_20260702_125026.log filter=lfs diff=lfs merge=lfs -text
 
 
 
 
58
  TimewarpVAE/RKDv3_DA_APHead_FM_RKD_CosineDecay_from_GR00TBase_to_120k/default/fm_rkdv3_timewarp_vae_da_aphead_cosine_decay_120k/checkpoint-90000/trainer_state.json filter=lfs diff=lfs merge=lfs -text
59
  TimewarpVAE/RKDv3_DA_APHead_FM_RKD_CosineDecay_from_GR00TBase_to_120k/default/fm_rkdv3_timewarp_vae_da_aphead_cosine_decay_120k/trainer_state.json filter=lfs diff=lfs merge=lfs -text
60
  TimewarpVAE/logs/RKDv3_TimewarpVAE_DA_APHead_FM_RKD_CosineDecay_from_GR00TBase_to_120k_gpu7_bs128_numgpu1_wandb_20260702_125026.log filter=lfs diff=lfs merge=lfs -text
61
+ RKDv3/base120_p2-60k_p3-60k_fm1-rkd0p5_atomic65_human/default/phase3/trainer_state.json filter=lfs diff=lfs merge=lfs -text
62
+ RKDv3/base120_p2-40k_p3-80k_fm1-rkd0p5_atomic65_human/default/phase3/trainer_state.json filter=lfs diff=lfs merge=lfs -text
63
+ RKDv3/base120_p2-60k_p3-60k_fm1-rkd0p5_atomic65_human/default/phase2/trainer_state.json filter=lfs diff=lfs merge=lfs -text
RKDv3/base120_p2-40k_p3-80k_fm1-rkd0p5_atomic65_human.stopped_nospace_20260708_204923/default/base120_p2-40k_p3-80k_fm1-rkd0p5_atomic65_human.yaml ADDED
@@ -0,0 +1,49 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ experiment_name: RKDv3 TimewarpVAE DA, Student:APHead from GR00T base to 120k on Atomic65
2
+ human, phase2 40k + phase3 80k, FM:RKD 1:0.5
3
+ policy_type: groot_RKDv3
4
+ base_model_path: nvidia/GR00T-N1.5-3B
5
+ output_root: ${HOME}/groot_robocasa/robocasa_v2/RKDv3/base120_p2-40k_p3-80k_fm1-rkd0p5_atomic65_human
6
+ dataset:
7
+ dataset_soup: atomic_pretrain_human
8
+ training:
9
+ num_gpus: 1
10
+ batch_size: 128
11
+ seed: 42
12
+ model:
13
+ rkd_teacher_source: timewarp_action
14
+ rkd_timewarp_vae_checkpoint: ${HOME}/groot_robocasa/TimeWarping/TimewarpVAE/robocasa_vanilla_t16_z8_g8_h32_reg0p01_beta0p01/checkpoint-step-60000-epoch-17.46
15
+ rkd_relation_mode: flatten
16
+ rkd_loss_type: distance_angle
17
+ rkd_student_reconstruction_head_pooling: attention
18
+ rkd_student_reconstruction_head_enabled: true
19
+ rkd_student_reconstruction_head_dim: 512
20
+ rkd_student_reconstruction_head_hidden_dim: 512
21
+ rkd_distance_loss_weight: 1.0
22
+ rkd_angle_loss_weight: 2.0
23
+ rkd_exclude_diagonal: true
24
+ phases:
25
+ - name: phase2_rkdv3_timewarp_vae_da_aphead_only_40k
26
+ max_steps: 40000
27
+ save_steps: 0
28
+ trainable:
29
+ preset: processing_line_only
30
+ tune_llm: false
31
+ tune_visual: false
32
+ tune_projector: false
33
+ tune_diffusion_model: false
34
+ losses:
35
+ rkd_enabled: true
36
+ rkd_fm_loss_weight: 0.0
37
+ rkd_loss_weight: 1.0
38
+ - name: phase3_fm_rkdv3_timewarp_vae_da_aphead_fixed_0p5_80k
39
+ max_steps: 80000
40
+ save_steps: 0
41
+ trainable:
42
+ tune_llm: false
43
+ tune_visual: false
44
+ tune_projector: true
45
+ tune_diffusion_model: true
46
+ losses:
47
+ rkd_enabled: true
48
+ rkd_fm_loss_weight: 1.0
49
+ rkd_loss_weight: 0.5
RKDv3/base120_p2-40k_p3-80k_fm1-rkd0p5_atomic65_human.stopped_nospace_20260708_204923/default/phase2/experiment_cfg/metadata.json ADDED
@@ -0,0 +1,431 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "new_embodiment": {
3
+ "statistics": {
4
+ "state": {
5
+ "base_position": {
6
+ "max": [
7
+ 7.778976917266846,
8
+ 0.4876587688922882,
9
+ 0.7223060727119446
10
+ ],
11
+ "min": [
12
+ -4.94293737411499,
13
+ -6.890198230743408,
14
+ 0.6994727253913879
15
+ ],
16
+ "mean": [
17
+ 2.645730005800653,
18
+ -2.110066447108802,
19
+ 0.7004254938507055
20
+ ],
21
+ "std": [
22
+ 1.674798582069234,
23
+ 1.350392482318102,
24
+ 0.001107956680671985
25
+ ],
26
+ "q01": [
27
+ -1.0132281893317567,
28
+ -5.366714195503661,
29
+ 0.699973412653353
30
+ ],
31
+ "q99": [
32
+ 6.099414437993715,
33
+ -0.5498886012762193,
34
+ 0.7032860957123939
35
+ ]
36
+ },
37
+ "base_rotation": {
38
+ "max": [
39
+ 0.0,
40
+ 0.0,
41
+ 1.0,
42
+ 1.0
43
+ ],
44
+ "min": [
45
+ 0.0,
46
+ 0.0,
47
+ -1.0,
48
+ 0.0
49
+ ],
50
+ "mean": [
51
+ 0.0,
52
+ 0.0,
53
+ 0.24103797327895668,
54
+ 0.6071393324239354
55
+ ],
56
+ "std": [
57
+ 0.0,
58
+ 0.0,
59
+ 0.6615962784298529,
60
+ 0.36814045893518177
61
+ ],
62
+ "q01": [
63
+ 0.0,
64
+ 0.0,
65
+ -1.0,
66
+ 3.3913522062507517e-06
67
+ ],
68
+ "q99": [
69
+ 0.0,
70
+ 0.0,
71
+ 1.0,
72
+ 1.0
73
+ ]
74
+ },
75
+ "end_effector_position_relative": {
76
+ "max": [
77
+ 0.9014528393745422,
78
+ 0.8569459915161133,
79
+ 1.053884744644165
80
+ ],
81
+ "min": [
82
+ -0.5828549861907959,
83
+ -0.8486236333847046,
84
+ -0.3446122705936432
85
+ ],
86
+ "mean": [
87
+ 0.2641355221895152,
88
+ -0.01768191244102862,
89
+ 0.4458847037294584
90
+ ],
91
+ "std": [
92
+ 0.16437879912713016,
93
+ 0.24835923712456218,
94
+ 0.2124190803465676
95
+ ],
96
+ "q01": [
97
+ -0.004962867024380738,
98
+ -0.46862868595225443,
99
+ 0.20316236863116172
100
+ ],
101
+ "q99": [
102
+ 0.5495510296644974,
103
+ 0.44201455499146414,
104
+ 0.6984316542098391
105
+ ]
106
+ },
107
+ "end_effector_rotation_relative": {
108
+ "max": [
109
+ 0.9999998807907104,
110
+ 0.9997354745864868,
111
+ 0.9900747537612915,
112
+ 0.9735639095306396
113
+ ],
114
+ "min": [
115
+ -0.9999930262565613,
116
+ -0.9999884366989136,
117
+ -0.9975369572639465,
118
+ 6.754255821306288e-08
119
+ ],
120
+ "mean": [
121
+ -0.23536101367097556,
122
+ 0.0384247508727208,
123
+ -0.06854891687606551,
124
+ 0.1400254132624415
125
+ ],
126
+ "std": [
127
+ 0.8296216819146821,
128
+ 0.30067208350163366,
129
+ 0.33390768096919515,
130
+ 0.16926885514344236
131
+ ],
132
+ "q01": [
133
+ -0.9950540525525874,
134
+ -0.5590028004550295,
135
+ -0.5635208777456633,
136
+ 0.001747048046951923
137
+ ],
138
+ "q99": [
139
+ 0.9472459325196646,
140
+ 0.6183763159959191,
141
+ 0.5108060516185887,
142
+ 0.46646525101830405
143
+ ]
144
+ },
145
+ "gripper_qpos": {
146
+ "max": [
147
+ 0.055869169533252716,
148
+ 0.010916369967162609
149
+ ],
150
+ "min": [
151
+ -0.011669117026031017,
152
+ -0.06307756155729294
153
+ ],
154
+ "mean": [
155
+ 0.031849369408926076,
156
+ -0.03200748595341724
157
+ ],
158
+ "std": [
159
+ 0.012248040928874927,
160
+ 0.01206276030599065
161
+ ],
162
+ "q01": [
163
+ 0.006329614505051435,
164
+ -0.0405872199089872
165
+ ],
166
+ "q99": [
167
+ 0.040445365520021676,
168
+ -0.006282430560590391
169
+ ]
170
+ }
171
+ },
172
+ "action": {
173
+ "base_motion": {
174
+ "max": [
175
+ 1.0,
176
+ 1.0,
177
+ 1.0,
178
+ 0.0
179
+ ],
180
+ "min": [
181
+ -1.0,
182
+ -1.0,
183
+ -1.0,
184
+ 0.0
185
+ ],
186
+ "mean": [
187
+ 0.0071452133647496785,
188
+ 0.0015603604277540566,
189
+ -0.0014052511062474704,
190
+ 0.0
191
+ ],
192
+ "std": [
193
+ 0.11471999232420287,
194
+ 0.10209228015027566,
195
+ 0.0855718789814838,
196
+ 0.0
197
+ ],
198
+ "q01": [
199
+ -0.13329522793794643,
200
+ -0.16270717101832055,
201
+ -0.12619798329427676,
202
+ 0.0
203
+ ],
204
+ "q99": [
205
+ 0.1721366453818688,
206
+ 0.1972873452547982,
207
+ 0.11128571706089413,
208
+ 0.0
209
+ ]
210
+ },
211
+ "control_mode": {
212
+ "max": [
213
+ 1.0
214
+ ],
215
+ "min": [
216
+ -1.0
217
+ ],
218
+ "mean": [
219
+ -0.9218734936549906
220
+ ],
221
+ "std": [
222
+ 0.3874975703570591
223
+ ],
224
+ "q01": [
225
+ -1.0
226
+ ],
227
+ "q99": [
228
+ -0.26692269247002653
229
+ ]
230
+ },
231
+ "end_effector_position": {
232
+ "max": [
233
+ 1.0,
234
+ 1.0,
235
+ 1.0
236
+ ],
237
+ "min": [
238
+ -1.0,
239
+ -1.0,
240
+ -1.0
241
+ ],
242
+ "mean": [
243
+ 0.002920238070268553,
244
+ -0.0060913649045883315,
245
+ -0.07322447393298565
246
+ ],
247
+ "std": [
248
+ 0.4482114094812734,
249
+ 0.41233706159751987,
250
+ 0.4006084454909327
251
+ ],
252
+ "q01": [
253
+ -0.934104737554819,
254
+ -0.9255792016234818,
255
+ -0.8085822192859831
256
+ ],
257
+ "q99": [
258
+ 0.7874185230421324,
259
+ 0.9004049161564378,
260
+ 0.8018332495119154
261
+ ]
262
+ },
263
+ "end_effector_rotation": {
264
+ "max": [
265
+ 1.0,
266
+ 1.0,
267
+ 1.0
268
+ ],
269
+ "min": [
270
+ -1.0,
271
+ -1.0,
272
+ -1.0
273
+ ],
274
+ "mean": [
275
+ 0.0077772595328458915,
276
+ -0.022620543966259083,
277
+ -0.00029332176634943376
278
+ ],
279
+ "std": [
280
+ 0.11155576533045976,
281
+ 0.13202104364511658,
282
+ 0.1240098754411023
283
+ ],
284
+ "q01": [
285
+ -0.2726786325766535,
286
+ -0.4125887853047325,
287
+ -0.3445183934868137
288
+ ],
289
+ "q99": [
290
+ 0.31803481867756295,
291
+ 0.3015153519871841,
292
+ 0.34730986869917346
293
+ ]
294
+ },
295
+ "gripper_close": {
296
+ "max": [
297
+ 1.0
298
+ ],
299
+ "min": [
300
+ -1.0
301
+ ],
302
+ "mean": [
303
+ -0.33422761717726496
304
+ ],
305
+ "std": [
306
+ 0.9425001619711008
307
+ ],
308
+ "q01": [
309
+ -1.0
310
+ ],
311
+ "q99": [
312
+ 0.681821585371573
313
+ ]
314
+ }
315
+ }
316
+ },
317
+ "modalities": {
318
+ "video": {
319
+ "robot0_eye_in_hand": {
320
+ "resolution": [
321
+ 256,
322
+ 256
323
+ ],
324
+ "channels": 3,
325
+ "fps": 20.0
326
+ },
327
+ "robot0_agentview_left": {
328
+ "resolution": [
329
+ 256,
330
+ 256
331
+ ],
332
+ "channels": 3,
333
+ "fps": 20.0
334
+ },
335
+ "robot0_agentview_right": {
336
+ "resolution": [
337
+ 256,
338
+ 256
339
+ ],
340
+ "channels": 3,
341
+ "fps": 20.0
342
+ }
343
+ },
344
+ "state": {
345
+ "base_position": {
346
+ "absolute": true,
347
+ "rotation_type": null,
348
+ "shape": [
349
+ 3
350
+ ],
351
+ "continuous": true
352
+ },
353
+ "base_rotation": {
354
+ "absolute": true,
355
+ "rotation_type": "quaternion",
356
+ "shape": [
357
+ 4
358
+ ],
359
+ "continuous": true
360
+ },
361
+ "end_effector_position_relative": {
362
+ "absolute": true,
363
+ "rotation_type": null,
364
+ "shape": [
365
+ 3
366
+ ],
367
+ "continuous": true
368
+ },
369
+ "end_effector_rotation_relative": {
370
+ "absolute": true,
371
+ "rotation_type": "quaternion",
372
+ "shape": [
373
+ 4
374
+ ],
375
+ "continuous": true
376
+ },
377
+ "gripper_qpos": {
378
+ "absolute": true,
379
+ "rotation_type": null,
380
+ "shape": [
381
+ 2
382
+ ],
383
+ "continuous": true
384
+ }
385
+ },
386
+ "action": {
387
+ "base_motion": {
388
+ "absolute": true,
389
+ "rotation_type": null,
390
+ "shape": [
391
+ 4
392
+ ],
393
+ "continuous": true
394
+ },
395
+ "control_mode": {
396
+ "absolute": true,
397
+ "rotation_type": null,
398
+ "shape": [
399
+ 1
400
+ ],
401
+ "continuous": true
402
+ },
403
+ "end_effector_position": {
404
+ "absolute": true,
405
+ "rotation_type": null,
406
+ "shape": [
407
+ 3
408
+ ],
409
+ "continuous": true
410
+ },
411
+ "end_effector_rotation": {
412
+ "absolute": true,
413
+ "rotation_type": "axis_angle",
414
+ "shape": [
415
+ 3
416
+ ],
417
+ "continuous": true
418
+ },
419
+ "gripper_close": {
420
+ "absolute": true,
421
+ "rotation_type": null,
422
+ "shape": [
423
+ 1
424
+ ],
425
+ "continuous": true
426
+ }
427
+ }
428
+ },
429
+ "embodiment_tag": "new_embodiment"
430
+ }
431
+ }
RKDv3/base120_p2-40k_p3-80k_fm1-rkd0p5_atomic65_human.stopped_nospace_20260708_204923/default/resolved_config.yaml ADDED
@@ -0,0 +1,50 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ experiment_name: RKDv3 TimewarpVAE DA, Student:APHead from GR00T base to 120k on Atomic65
2
+ human, phase2 40k + phase3 80k, FM:RKD 1:0.5
3
+ policy_type: groot_RKDv3
4
+ base_model_path: nvidia/GR00T-N1.5-3B
5
+ output_root: ${HOME}/groot_robocasa/robocasa_v2/RKDv3/base120_p2-40k_p3-80k_fm1-rkd0p5_atomic65_human
6
+ dataset:
7
+ dataset_soup: atomic_pretrain_human
8
+ training:
9
+ num_gpus: 1
10
+ batch_size: 128
11
+ seed: 42
12
+ model:
13
+ rkd_teacher_source: timewarp_action
14
+ rkd_timewarp_vae_checkpoint: ${HOME}/groot_robocasa/TimeWarping/TimewarpVAE/robocasa_vanilla_t16_z8_g8_h32_reg0p01_beta0p01/checkpoint-step-60000-epoch-17.46
15
+ rkd_relation_mode: flatten
16
+ rkd_loss_type: distance_angle
17
+ rkd_student_reconstruction_head_pooling: attention
18
+ rkd_student_reconstruction_head_enabled: true
19
+ rkd_student_reconstruction_head_dim: 512
20
+ rkd_student_reconstruction_head_hidden_dim: 512
21
+ rkd_distance_loss_weight: 1.0
22
+ rkd_angle_loss_weight: 2.0
23
+ rkd_exclude_diagonal: true
24
+ phases:
25
+ - name: phase2_rkdv3_timewarp_vae_da_aphead_only_40k
26
+ max_steps: 40000
27
+ save_steps: 0
28
+ trainable:
29
+ preset: processing_line_only
30
+ tune_llm: false
31
+ tune_visual: false
32
+ tune_projector: false
33
+ tune_diffusion_model: false
34
+ losses:
35
+ rkd_enabled: true
36
+ rkd_fm_loss_weight: 0.0
37
+ rkd_loss_weight: 1.0
38
+ - name: phase3_fm_rkdv3_timewarp_vae_da_aphead_fixed_0p5_80k
39
+ max_steps: 80000
40
+ save_steps: 0
41
+ trainable:
42
+ tune_llm: false
43
+ tune_visual: false
44
+ tune_projector: true
45
+ tune_diffusion_model: true
46
+ losses:
47
+ rkd_enabled: true
48
+ rkd_fm_loss_weight: 1.0
49
+ rkd_loss_weight: 0.5
50
+ resolved_sweep: {}
RKDv3/base120_p2-40k_p3-80k_fm1-rkd0p5_atomic65_human/default/base120_p2-40k_p3-80k_fm1-rkd0p5_atomic65_human.yaml ADDED
@@ -0,0 +1,49 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ experiment_name: RKDv3 TimewarpVAE DA, Student:APHead from GR00T base to 120k on Atomic65
2
+ human, phase2 40k + phase3 80k, FM:RKD 1:0.5
3
+ policy_type: groot_RKDv3
4
+ base_model_path: nvidia/GR00T-N1.5-3B
5
+ output_root: ${HOME}/groot_robocasa/robocasa_v2/RKDv3/base120_p2-40k_p3-80k_fm1-rkd0p5_atomic65_human
6
+ dataset:
7
+ dataset_soup: atomic_pretrain_human
8
+ training:
9
+ num_gpus: 1
10
+ batch_size: 128
11
+ seed: 42
12
+ model:
13
+ rkd_teacher_source: timewarp_action
14
+ rkd_timewarp_vae_checkpoint: ${HOME}/groot_robocasa/TimeWarping/TimewarpVAE/atomic65_z8_g16_h32_tw32_reg0p01_gb_ctrl_30k/checkpoint-step-30000-epoch-2.73
15
+ rkd_relation_mode: flatten
16
+ rkd_loss_type: distance_angle
17
+ rkd_student_reconstruction_head_pooling: attention
18
+ rkd_student_reconstruction_head_enabled: true
19
+ rkd_student_reconstruction_head_dim: 512
20
+ rkd_student_reconstruction_head_hidden_dim: 512
21
+ rkd_distance_loss_weight: 1.0
22
+ rkd_angle_loss_weight: 2.0
23
+ rkd_exclude_diagonal: true
24
+ phases:
25
+ - name: phase2_rkdv3_timewarp_vae_da_aphead_only_40k
26
+ max_steps: 40000
27
+ save_steps: 0
28
+ trainable:
29
+ preset: processing_line_only
30
+ tune_llm: false
31
+ tune_visual: false
32
+ tune_projector: false
33
+ tune_diffusion_model: false
34
+ losses:
35
+ rkd_enabled: true
36
+ rkd_fm_loss_weight: 0.0
37
+ rkd_loss_weight: 1.0
38
+ - name: phase3_fm_rkdv3_timewarp_vae_da_aphead_fixed_0p5_80k
39
+ max_steps: 80000
40
+ save_steps: 0
41
+ trainable:
42
+ tune_llm: false
43
+ tune_visual: false
44
+ tune_projector: true
45
+ tune_diffusion_model: true
46
+ losses:
47
+ rkd_enabled: true
48
+ rkd_fm_loss_weight: 1.0
49
+ rkd_loss_weight: 0.5
RKDv3/base120_p2-40k_p3-80k_fm1-rkd0p5_atomic65_human/default/phase2/config.json ADDED
@@ -0,0 +1,101 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "action_dim": 32,
3
+ "action_head_cfg": {
4
+ "action_dim": 32,
5
+ "action_horizon": 16,
6
+ "add_pos_embed": true,
7
+ "backbone_embedding_dim": 2048,
8
+ "diffusion_model_cfg": {
9
+ "attention_head_dim": 48,
10
+ "cross_attention_dim": 2048,
11
+ "dropout": 0.2,
12
+ "final_dropout": true,
13
+ "interleave_self_attention": true,
14
+ "norm_type": "ada_norm",
15
+ "num_attention_heads": 32,
16
+ "num_layers": 16,
17
+ "output_dim": 1024,
18
+ "positional_embeddings": null
19
+ },
20
+ "hidden_size": 1024,
21
+ "input_embedding_dim": 1536,
22
+ "max_action_dim": 32,
23
+ "max_state_dim": 64,
24
+ "model_dtype": "float32",
25
+ "noise_beta_alpha": 1.5,
26
+ "noise_beta_beta": 1.0,
27
+ "noise_s": 0.999,
28
+ "num_inference_timesteps": 4,
29
+ "num_target_vision_tokens": 32,
30
+ "num_timestep_buckets": 1000,
31
+ "tune_diffusion_model": true,
32
+ "tune_projector": true,
33
+ "use_vlln": true,
34
+ "vl_self_attention_cfg": {
35
+ "attention_head_dim": 64,
36
+ "dropout": 0.2,
37
+ "final_dropout": true,
38
+ "num_attention_heads": 32,
39
+ "num_layers": 4,
40
+ "positional_embeddings": null
41
+ }
42
+ },
43
+ "action_horizon": 16,
44
+ "architectures": [
45
+ "GR00T_N1_5_RKD"
46
+ ],
47
+ "attn_implementation": null,
48
+ "backbone_cfg": {
49
+ "eagle_path": "NVEagle/eagle_er-qwen3_1_7B-Siglip2_400M_stage1_5_128gpu_er_v7_1mlp_nops",
50
+ "load_bf16": false,
51
+ "project_to_dim": null,
52
+ "reproject_vision": false,
53
+ "select_layer": 12,
54
+ "tune_llm": false,
55
+ "tune_visual": true,
56
+ "use_flash_attention": true
57
+ },
58
+ "compute_dtype": "bfloat16",
59
+ "hidden_size": 2048,
60
+ "model_dtype": "float32",
61
+ "model_type": "gr00t_n1_5",
62
+ "rkd_action_encoder_cls_ffn_hidden_dim": 2048,
63
+ "rkd_action_encoder_cls_num_heads": 8,
64
+ "rkd_action_encoder_cls_num_layers": 1,
65
+ "rkd_action_encoder_projector_dim": 512,
66
+ "rkd_action_encoder_projector_enabled": false,
67
+ "rkd_action_encoder_projector_pooling": "flatten",
68
+ "rkd_action_temp": 0.1,
69
+ "rkd_angle_loss_weight": 2.0,
70
+ "rkd_distance_loss_weight": 1.0,
71
+ "rkd_dtw_adaptive_floor": 0.05,
72
+ "rkd_dtw_adaptive_mix_lambda": 0.5,
73
+ "rkd_dtw_base_weight": 1.0,
74
+ "rkd_dtw_gripper_weight": 0.5,
75
+ "rkd_dtw_mix_mode": "distance_mix",
76
+ "rkd_dtw_normalize_components": true,
77
+ "rkd_dtw_pos_weight": 1.0,
78
+ "rkd_dtw_rot_weight": 1.0,
79
+ "rkd_dtw_timing_log_interval": 1000,
80
+ "rkd_enabled": true,
81
+ "rkd_exclude_diagonal": true,
82
+ "rkd_fm_loss_weight": 0.0,
83
+ "rkd_loss_type": "distance_angle",
84
+ "rkd_loss_weight": 1.0,
85
+ "rkd_loss_weight_end": 0.0,
86
+ "rkd_loss_weight_schedule": null,
87
+ "rkd_loss_weight_start": 0.1,
88
+ "rkd_raw_action_projector_dim": 512,
89
+ "rkd_raw_action_projector_enabled": false,
90
+ "rkd_raw_action_projector_pooling": "flatten",
91
+ "rkd_relation_mode": "flatten",
92
+ "rkd_student_reconstruction_head_dim": 512,
93
+ "rkd_student_reconstruction_head_enabled": true,
94
+ "rkd_student_reconstruction_head_hidden_dim": 512,
95
+ "rkd_student_reconstruction_head_pooling": "attention",
96
+ "rkd_teacher_source": "timewarp_action",
97
+ "rkd_timewarp_vae_checkpoint": "${HOME}/groot_robocasa/TimeWarping/TimewarpVAE/atomic65_z8_g16_h32_tw32_reg0p01_gb_ctrl_30k/checkpoint-step-30000-epoch-2.73",
98
+ "rkd_vlm_temp": 0.1,
99
+ "torch_dtype": "bfloat16",
100
+ "transformers_version": "4.51.3"
101
+ }
RKDv3/base120_p2-40k_p3-80k_fm1-rkd0p5_atomic65_human/default/phase2/experiment_cfg/metadata.json ADDED
@@ -0,0 +1,431 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "new_embodiment": {
3
+ "statistics": {
4
+ "state": {
5
+ "base_position": {
6
+ "max": [
7
+ 7.778976917266846,
8
+ 0.4876587688922882,
9
+ 0.7223060727119446
10
+ ],
11
+ "min": [
12
+ -4.94293737411499,
13
+ -6.890198230743408,
14
+ 0.6994727253913879
15
+ ],
16
+ "mean": [
17
+ 2.645730005800653,
18
+ -2.110066447108802,
19
+ 0.7004254938507055
20
+ ],
21
+ "std": [
22
+ 1.674798582069234,
23
+ 1.350392482318102,
24
+ 0.001107956680671985
25
+ ],
26
+ "q01": [
27
+ -1.0132281893317567,
28
+ -5.366714195503661,
29
+ 0.699973412653353
30
+ ],
31
+ "q99": [
32
+ 6.099414437993715,
33
+ -0.5498886012762193,
34
+ 0.7032860957123939
35
+ ]
36
+ },
37
+ "base_rotation": {
38
+ "max": [
39
+ 0.0,
40
+ 0.0,
41
+ 1.0,
42
+ 1.0
43
+ ],
44
+ "min": [
45
+ 0.0,
46
+ 0.0,
47
+ -1.0,
48
+ 0.0
49
+ ],
50
+ "mean": [
51
+ 0.0,
52
+ 0.0,
53
+ 0.24103797327895668,
54
+ 0.6071393324239354
55
+ ],
56
+ "std": [
57
+ 0.0,
58
+ 0.0,
59
+ 0.6615962784298529,
60
+ 0.36814045893518177
61
+ ],
62
+ "q01": [
63
+ 0.0,
64
+ 0.0,
65
+ -1.0,
66
+ 3.3913522062507517e-06
67
+ ],
68
+ "q99": [
69
+ 0.0,
70
+ 0.0,
71
+ 1.0,
72
+ 1.0
73
+ ]
74
+ },
75
+ "end_effector_position_relative": {
76
+ "max": [
77
+ 0.9014528393745422,
78
+ 0.8569459915161133,
79
+ 1.053884744644165
80
+ ],
81
+ "min": [
82
+ -0.5828549861907959,
83
+ -0.8486236333847046,
84
+ -0.3446122705936432
85
+ ],
86
+ "mean": [
87
+ 0.2641355221895152,
88
+ -0.01768191244102862,
89
+ 0.4458847037294584
90
+ ],
91
+ "std": [
92
+ 0.16437879912713016,
93
+ 0.24835923712456218,
94
+ 0.2124190803465676
95
+ ],
96
+ "q01": [
97
+ -0.004962867024380738,
98
+ -0.46862868595225443,
99
+ 0.20316236863116172
100
+ ],
101
+ "q99": [
102
+ 0.5495510296644974,
103
+ 0.44201455499146414,
104
+ 0.6984316542098391
105
+ ]
106
+ },
107
+ "end_effector_rotation_relative": {
108
+ "max": [
109
+ 0.9999998807907104,
110
+ 0.9997354745864868,
111
+ 0.9900747537612915,
112
+ 0.9735639095306396
113
+ ],
114
+ "min": [
115
+ -0.9999930262565613,
116
+ -0.9999884366989136,
117
+ -0.9975369572639465,
118
+ 6.754255821306288e-08
119
+ ],
120
+ "mean": [
121
+ -0.23536101367097556,
122
+ 0.0384247508727208,
123
+ -0.06854891687606551,
124
+ 0.1400254132624415
125
+ ],
126
+ "std": [
127
+ 0.8296216819146821,
128
+ 0.30067208350163366,
129
+ 0.33390768096919515,
130
+ 0.16926885514344236
131
+ ],
132
+ "q01": [
133
+ -0.9950540525525874,
134
+ -0.5590028004550295,
135
+ -0.5635208777456633,
136
+ 0.001747048046951923
137
+ ],
138
+ "q99": [
139
+ 0.9472459325196646,
140
+ 0.6183763159959191,
141
+ 0.5108060516185887,
142
+ 0.46646525101830405
143
+ ]
144
+ },
145
+ "gripper_qpos": {
146
+ "max": [
147
+ 0.055869169533252716,
148
+ 0.010916369967162609
149
+ ],
150
+ "min": [
151
+ -0.011669117026031017,
152
+ -0.06307756155729294
153
+ ],
154
+ "mean": [
155
+ 0.031849369408926076,
156
+ -0.03200748595341724
157
+ ],
158
+ "std": [
159
+ 0.012248040928874927,
160
+ 0.01206276030599065
161
+ ],
162
+ "q01": [
163
+ 0.006329614505051435,
164
+ -0.0405872199089872
165
+ ],
166
+ "q99": [
167
+ 0.040445365520021676,
168
+ -0.006282430560590391
169
+ ]
170
+ }
171
+ },
172
+ "action": {
173
+ "base_motion": {
174
+ "max": [
175
+ 1.0,
176
+ 1.0,
177
+ 1.0,
178
+ 0.0
179
+ ],
180
+ "min": [
181
+ -1.0,
182
+ -1.0,
183
+ -1.0,
184
+ 0.0
185
+ ],
186
+ "mean": [
187
+ 0.0071452133647496785,
188
+ 0.0015603604277540566,
189
+ -0.0014052511062474704,
190
+ 0.0
191
+ ],
192
+ "std": [
193
+ 0.11471999232420287,
194
+ 0.10209228015027566,
195
+ 0.0855718789814838,
196
+ 0.0
197
+ ],
198
+ "q01": [
199
+ -0.13329522793794643,
200
+ -0.16270717101832055,
201
+ -0.12619798329427676,
202
+ 0.0
203
+ ],
204
+ "q99": [
205
+ 0.1721366453818688,
206
+ 0.1972873452547982,
207
+ 0.11128571706089413,
208
+ 0.0
209
+ ]
210
+ },
211
+ "control_mode": {
212
+ "max": [
213
+ 1.0
214
+ ],
215
+ "min": [
216
+ -1.0
217
+ ],
218
+ "mean": [
219
+ -0.9218734936549906
220
+ ],
221
+ "std": [
222
+ 0.3874975703570591
223
+ ],
224
+ "q01": [
225
+ -1.0
226
+ ],
227
+ "q99": [
228
+ -0.26692269247002653
229
+ ]
230
+ },
231
+ "end_effector_position": {
232
+ "max": [
233
+ 1.0,
234
+ 1.0,
235
+ 1.0
236
+ ],
237
+ "min": [
238
+ -1.0,
239
+ -1.0,
240
+ -1.0
241
+ ],
242
+ "mean": [
243
+ 0.002920238070268553,
244
+ -0.0060913649045883315,
245
+ -0.07322447393298565
246
+ ],
247
+ "std": [
248
+ 0.4482114094812734,
249
+ 0.41233706159751987,
250
+ 0.4006084454909327
251
+ ],
252
+ "q01": [
253
+ -0.934104737554819,
254
+ -0.9255792016234818,
255
+ -0.8085822192859831
256
+ ],
257
+ "q99": [
258
+ 0.7874185230421324,
259
+ 0.9004049161564378,
260
+ 0.8018332495119154
261
+ ]
262
+ },
263
+ "end_effector_rotation": {
264
+ "max": [
265
+ 1.0,
266
+ 1.0,
267
+ 1.0
268
+ ],
269
+ "min": [
270
+ -1.0,
271
+ -1.0,
272
+ -1.0
273
+ ],
274
+ "mean": [
275
+ 0.0077772595328458915,
276
+ -0.022620543966259083,
277
+ -0.00029332176634943376
278
+ ],
279
+ "std": [
280
+ 0.11155576533045976,
281
+ 0.13202104364511658,
282
+ 0.1240098754411023
283
+ ],
284
+ "q01": [
285
+ -0.2726786325766535,
286
+ -0.4125887853047325,
287
+ -0.3445183934868137
288
+ ],
289
+ "q99": [
290
+ 0.31803481867756295,
291
+ 0.3015153519871841,
292
+ 0.34730986869917346
293
+ ]
294
+ },
295
+ "gripper_close": {
296
+ "max": [
297
+ 1.0
298
+ ],
299
+ "min": [
300
+ -1.0
301
+ ],
302
+ "mean": [
303
+ -0.33422761717726496
304
+ ],
305
+ "std": [
306
+ 0.9425001619711008
307
+ ],
308
+ "q01": [
309
+ -1.0
310
+ ],
311
+ "q99": [
312
+ 0.681821585371573
313
+ ]
314
+ }
315
+ }
316
+ },
317
+ "modalities": {
318
+ "video": {
319
+ "robot0_eye_in_hand": {
320
+ "resolution": [
321
+ 256,
322
+ 256
323
+ ],
324
+ "channels": 3,
325
+ "fps": 20.0
326
+ },
327
+ "robot0_agentview_left": {
328
+ "resolution": [
329
+ 256,
330
+ 256
331
+ ],
332
+ "channels": 3,
333
+ "fps": 20.0
334
+ },
335
+ "robot0_agentview_right": {
336
+ "resolution": [
337
+ 256,
338
+ 256
339
+ ],
340
+ "channels": 3,
341
+ "fps": 20.0
342
+ }
343
+ },
344
+ "state": {
345
+ "base_position": {
346
+ "absolute": true,
347
+ "rotation_type": null,
348
+ "shape": [
349
+ 3
350
+ ],
351
+ "continuous": true
352
+ },
353
+ "base_rotation": {
354
+ "absolute": true,
355
+ "rotation_type": "quaternion",
356
+ "shape": [
357
+ 4
358
+ ],
359
+ "continuous": true
360
+ },
361
+ "end_effector_position_relative": {
362
+ "absolute": true,
363
+ "rotation_type": null,
364
+ "shape": [
365
+ 3
366
+ ],
367
+ "continuous": true
368
+ },
369
+ "end_effector_rotation_relative": {
370
+ "absolute": true,
371
+ "rotation_type": "quaternion",
372
+ "shape": [
373
+ 4
374
+ ],
375
+ "continuous": true
376
+ },
377
+ "gripper_qpos": {
378
+ "absolute": true,
379
+ "rotation_type": null,
380
+ "shape": [
381
+ 2
382
+ ],
383
+ "continuous": true
384
+ }
385
+ },
386
+ "action": {
387
+ "base_motion": {
388
+ "absolute": true,
389
+ "rotation_type": null,
390
+ "shape": [
391
+ 4
392
+ ],
393
+ "continuous": true
394
+ },
395
+ "control_mode": {
396
+ "absolute": true,
397
+ "rotation_type": null,
398
+ "shape": [
399
+ 1
400
+ ],
401
+ "continuous": true
402
+ },
403
+ "end_effector_position": {
404
+ "absolute": true,
405
+ "rotation_type": null,
406
+ "shape": [
407
+ 3
408
+ ],
409
+ "continuous": true
410
+ },
411
+ "end_effector_rotation": {
412
+ "absolute": true,
413
+ "rotation_type": "axis_angle",
414
+ "shape": [
415
+ 3
416
+ ],
417
+ "continuous": true
418
+ },
419
+ "gripper_close": {
420
+ "absolute": true,
421
+ "rotation_type": null,
422
+ "shape": [
423
+ 1
424
+ ],
425
+ "continuous": true
426
+ }
427
+ }
428
+ },
429
+ "embodiment_tag": "new_embodiment"
430
+ }
431
+ }
RKDv3/base120_p2-40k_p3-80k_fm1-rkd0p5_atomic65_human/default/phase2/model-00001-of-00002.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0e65895a7b6c052e58b679bdd68bff54a68694a81f38e23d3e03a92059b3600b
3
+ size 4999367032
RKDv3/base120_p2-40k_p3-80k_fm1-rkd0p5_atomic65_human/default/phase2/model-00002-of-00002.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:92af9ac4022f93629bcae242d1c22b5d2706bd79c0b5dba9a1676f10aeb77d9d
3
+ size 2664461700
RKDv3/base120_p2-40k_p3-80k_fm1-rkd0p5_atomic65_human/default/phase2/model.safetensors.index.json ADDED
The diff for this file is too large to render. See raw diff
 
RKDv3/base120_p2-40k_p3-80k_fm1-rkd0p5_atomic65_human/default/phase2/trainer_state.json ADDED
The diff for this file is too large to render. See raw diff
 
RKDv3/base120_p2-40k_p3-80k_fm1-rkd0p5_atomic65_human/default/phase2/training_args.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b12a8ec3c4989036079d41526de61c9f07a256aa1529292eb10871a19be9b256
3
+ size 6033
RKDv3/base120_p2-40k_p3-80k_fm1-rkd0p5_atomic65_human/default/phase3/config.json ADDED
@@ -0,0 +1,101 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "action_dim": 32,
3
+ "action_head_cfg": {
4
+ "action_dim": 32,
5
+ "action_horizon": 16,
6
+ "add_pos_embed": true,
7
+ "backbone_embedding_dim": 2048,
8
+ "diffusion_model_cfg": {
9
+ "attention_head_dim": 48,
10
+ "cross_attention_dim": 2048,
11
+ "dropout": 0.2,
12
+ "final_dropout": true,
13
+ "interleave_self_attention": true,
14
+ "norm_type": "ada_norm",
15
+ "num_attention_heads": 32,
16
+ "num_layers": 16,
17
+ "output_dim": 1024,
18
+ "positional_embeddings": null
19
+ },
20
+ "hidden_size": 1024,
21
+ "input_embedding_dim": 1536,
22
+ "max_action_dim": 32,
23
+ "max_state_dim": 64,
24
+ "model_dtype": "float32",
25
+ "noise_beta_alpha": 1.5,
26
+ "noise_beta_beta": 1.0,
27
+ "noise_s": 0.999,
28
+ "num_inference_timesteps": 4,
29
+ "num_target_vision_tokens": 32,
30
+ "num_timestep_buckets": 1000,
31
+ "tune_diffusion_model": true,
32
+ "tune_projector": true,
33
+ "use_vlln": true,
34
+ "vl_self_attention_cfg": {
35
+ "attention_head_dim": 64,
36
+ "dropout": 0.2,
37
+ "final_dropout": true,
38
+ "num_attention_heads": 32,
39
+ "num_layers": 4,
40
+ "positional_embeddings": null
41
+ }
42
+ },
43
+ "action_horizon": 16,
44
+ "architectures": [
45
+ "GR00T_N1_5_RKD"
46
+ ],
47
+ "attn_implementation": null,
48
+ "backbone_cfg": {
49
+ "eagle_path": "NVEagle/eagle_er-qwen3_1_7B-Siglip2_400M_stage1_5_128gpu_er_v7_1mlp_nops",
50
+ "load_bf16": false,
51
+ "project_to_dim": null,
52
+ "reproject_vision": false,
53
+ "select_layer": 12,
54
+ "tune_llm": false,
55
+ "tune_visual": true,
56
+ "use_flash_attention": true
57
+ },
58
+ "compute_dtype": "bfloat16",
59
+ "hidden_size": 2048,
60
+ "model_dtype": "float32",
61
+ "model_type": "gr00t_n1_5",
62
+ "rkd_action_encoder_cls_ffn_hidden_dim": 2048,
63
+ "rkd_action_encoder_cls_num_heads": 8,
64
+ "rkd_action_encoder_cls_num_layers": 1,
65
+ "rkd_action_encoder_projector_dim": 512,
66
+ "rkd_action_encoder_projector_enabled": false,
67
+ "rkd_action_encoder_projector_pooling": "flatten",
68
+ "rkd_action_temp": 0.1,
69
+ "rkd_angle_loss_weight": 2.0,
70
+ "rkd_distance_loss_weight": 1.0,
71
+ "rkd_dtw_adaptive_floor": 0.05,
72
+ "rkd_dtw_adaptive_mix_lambda": 0.5,
73
+ "rkd_dtw_base_weight": 1.0,
74
+ "rkd_dtw_gripper_weight": 0.5,
75
+ "rkd_dtw_mix_mode": "distance_mix",
76
+ "rkd_dtw_normalize_components": true,
77
+ "rkd_dtw_pos_weight": 1.0,
78
+ "rkd_dtw_rot_weight": 1.0,
79
+ "rkd_dtw_timing_log_interval": 1000,
80
+ "rkd_enabled": true,
81
+ "rkd_exclude_diagonal": true,
82
+ "rkd_fm_loss_weight": 1.0,
83
+ "rkd_loss_type": "distance_angle",
84
+ "rkd_loss_weight": 0.5,
85
+ "rkd_loss_weight_end": 0.0,
86
+ "rkd_loss_weight_schedule": null,
87
+ "rkd_loss_weight_start": 0.1,
88
+ "rkd_raw_action_projector_dim": 512,
89
+ "rkd_raw_action_projector_enabled": false,
90
+ "rkd_raw_action_projector_pooling": "flatten",
91
+ "rkd_relation_mode": "flatten",
92
+ "rkd_student_reconstruction_head_dim": 512,
93
+ "rkd_student_reconstruction_head_enabled": true,
94
+ "rkd_student_reconstruction_head_hidden_dim": 512,
95
+ "rkd_student_reconstruction_head_pooling": "attention",
96
+ "rkd_teacher_source": "timewarp_action",
97
+ "rkd_timewarp_vae_checkpoint": "${HOME}/groot_robocasa/TimeWarping/TimewarpVAE/atomic65_z8_g16_h32_tw32_reg0p01_gb_ctrl_30k/checkpoint-step-30000-epoch-2.73",
98
+ "rkd_vlm_temp": 0.1,
99
+ "torch_dtype": "bfloat16",
100
+ "transformers_version": "4.51.3"
101
+ }
RKDv3/base120_p2-40k_p3-80k_fm1-rkd0p5_atomic65_human/default/phase3/experiment_cfg/metadata.json ADDED
@@ -0,0 +1,431 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "new_embodiment": {
3
+ "statistics": {
4
+ "state": {
5
+ "base_position": {
6
+ "max": [
7
+ 7.778976917266846,
8
+ 0.4876587688922882,
9
+ 0.7223060727119446
10
+ ],
11
+ "min": [
12
+ -4.94293737411499,
13
+ -6.890198230743408,
14
+ 0.6994727253913879
15
+ ],
16
+ "mean": [
17
+ 2.645730005800653,
18
+ -2.110066447108802,
19
+ 0.7004254938507055
20
+ ],
21
+ "std": [
22
+ 1.674798582069234,
23
+ 1.350392482318102,
24
+ 0.001107956680671985
25
+ ],
26
+ "q01": [
27
+ -1.0132281893317567,
28
+ -5.366714195503661,
29
+ 0.699973412653353
30
+ ],
31
+ "q99": [
32
+ 6.099414437993715,
33
+ -0.5498886012762193,
34
+ 0.7032860957123939
35
+ ]
36
+ },
37
+ "base_rotation": {
38
+ "max": [
39
+ 0.0,
40
+ 0.0,
41
+ 1.0,
42
+ 1.0
43
+ ],
44
+ "min": [
45
+ 0.0,
46
+ 0.0,
47
+ -1.0,
48
+ 0.0
49
+ ],
50
+ "mean": [
51
+ 0.0,
52
+ 0.0,
53
+ 0.24103797327895668,
54
+ 0.6071393324239354
55
+ ],
56
+ "std": [
57
+ 0.0,
58
+ 0.0,
59
+ 0.6615962784298529,
60
+ 0.36814045893518177
61
+ ],
62
+ "q01": [
63
+ 0.0,
64
+ 0.0,
65
+ -1.0,
66
+ 3.3913522062507517e-06
67
+ ],
68
+ "q99": [
69
+ 0.0,
70
+ 0.0,
71
+ 1.0,
72
+ 1.0
73
+ ]
74
+ },
75
+ "end_effector_position_relative": {
76
+ "max": [
77
+ 0.9014528393745422,
78
+ 0.8569459915161133,
79
+ 1.053884744644165
80
+ ],
81
+ "min": [
82
+ -0.5828549861907959,
83
+ -0.8486236333847046,
84
+ -0.3446122705936432
85
+ ],
86
+ "mean": [
87
+ 0.2641355221895152,
88
+ -0.01768191244102862,
89
+ 0.4458847037294584
90
+ ],
91
+ "std": [
92
+ 0.16437879912713016,
93
+ 0.24835923712456218,
94
+ 0.2124190803465676
95
+ ],
96
+ "q01": [
97
+ -0.004962867024380738,
98
+ -0.46862868595225443,
99
+ 0.20316236863116172
100
+ ],
101
+ "q99": [
102
+ 0.5495510296644974,
103
+ 0.44201455499146414,
104
+ 0.6984316542098391
105
+ ]
106
+ },
107
+ "end_effector_rotation_relative": {
108
+ "max": [
109
+ 0.9999998807907104,
110
+ 0.9997354745864868,
111
+ 0.9900747537612915,
112
+ 0.9735639095306396
113
+ ],
114
+ "min": [
115
+ -0.9999930262565613,
116
+ -0.9999884366989136,
117
+ -0.9975369572639465,
118
+ 6.754255821306288e-08
119
+ ],
120
+ "mean": [
121
+ -0.23536101367097556,
122
+ 0.0384247508727208,
123
+ -0.06854891687606551,
124
+ 0.1400254132624415
125
+ ],
126
+ "std": [
127
+ 0.8296216819146821,
128
+ 0.30067208350163366,
129
+ 0.33390768096919515,
130
+ 0.16926885514344236
131
+ ],
132
+ "q01": [
133
+ -0.9950540525525874,
134
+ -0.5590028004550295,
135
+ -0.5635208777456633,
136
+ 0.001747048046951923
137
+ ],
138
+ "q99": [
139
+ 0.9472459325196646,
140
+ 0.6183763159959191,
141
+ 0.5108060516185887,
142
+ 0.46646525101830405
143
+ ]
144
+ },
145
+ "gripper_qpos": {
146
+ "max": [
147
+ 0.055869169533252716,
148
+ 0.010916369967162609
149
+ ],
150
+ "min": [
151
+ -0.011669117026031017,
152
+ -0.06307756155729294
153
+ ],
154
+ "mean": [
155
+ 0.031849369408926076,
156
+ -0.03200748595341724
157
+ ],
158
+ "std": [
159
+ 0.012248040928874927,
160
+ 0.01206276030599065
161
+ ],
162
+ "q01": [
163
+ 0.006329614505051435,
164
+ -0.0405872199089872
165
+ ],
166
+ "q99": [
167
+ 0.040445365520021676,
168
+ -0.006282430560590391
169
+ ]
170
+ }
171
+ },
172
+ "action": {
173
+ "base_motion": {
174
+ "max": [
175
+ 1.0,
176
+ 1.0,
177
+ 1.0,
178
+ 0.0
179
+ ],
180
+ "min": [
181
+ -1.0,
182
+ -1.0,
183
+ -1.0,
184
+ 0.0
185
+ ],
186
+ "mean": [
187
+ 0.0071452133647496785,
188
+ 0.0015603604277540566,
189
+ -0.0014052511062474704,
190
+ 0.0
191
+ ],
192
+ "std": [
193
+ 0.11471999232420287,
194
+ 0.10209228015027566,
195
+ 0.0855718789814838,
196
+ 0.0
197
+ ],
198
+ "q01": [
199
+ -0.13329522793794643,
200
+ -0.16270717101832055,
201
+ -0.12619798329427676,
202
+ 0.0
203
+ ],
204
+ "q99": [
205
+ 0.1721366453818688,
206
+ 0.1972873452547982,
207
+ 0.11128571706089413,
208
+ 0.0
209
+ ]
210
+ },
211
+ "control_mode": {
212
+ "max": [
213
+ 1.0
214
+ ],
215
+ "min": [
216
+ -1.0
217
+ ],
218
+ "mean": [
219
+ -0.9218734936549906
220
+ ],
221
+ "std": [
222
+ 0.3874975703570591
223
+ ],
224
+ "q01": [
225
+ -1.0
226
+ ],
227
+ "q99": [
228
+ -0.26692269247002653
229
+ ]
230
+ },
231
+ "end_effector_position": {
232
+ "max": [
233
+ 1.0,
234
+ 1.0,
235
+ 1.0
236
+ ],
237
+ "min": [
238
+ -1.0,
239
+ -1.0,
240
+ -1.0
241
+ ],
242
+ "mean": [
243
+ 0.002920238070268553,
244
+ -0.0060913649045883315,
245
+ -0.07322447393298565
246
+ ],
247
+ "std": [
248
+ 0.4482114094812734,
249
+ 0.41233706159751987,
250
+ 0.4006084454909327
251
+ ],
252
+ "q01": [
253
+ -0.934104737554819,
254
+ -0.9255792016234818,
255
+ -0.8085822192859831
256
+ ],
257
+ "q99": [
258
+ 0.7874185230421324,
259
+ 0.9004049161564378,
260
+ 0.8018332495119154
261
+ ]
262
+ },
263
+ "end_effector_rotation": {
264
+ "max": [
265
+ 1.0,
266
+ 1.0,
267
+ 1.0
268
+ ],
269
+ "min": [
270
+ -1.0,
271
+ -1.0,
272
+ -1.0
273
+ ],
274
+ "mean": [
275
+ 0.0077772595328458915,
276
+ -0.022620543966259083,
277
+ -0.00029332176634943376
278
+ ],
279
+ "std": [
280
+ 0.11155576533045976,
281
+ 0.13202104364511658,
282
+ 0.1240098754411023
283
+ ],
284
+ "q01": [
285
+ -0.2726786325766535,
286
+ -0.4125887853047325,
287
+ -0.3445183934868137
288
+ ],
289
+ "q99": [
290
+ 0.31803481867756295,
291
+ 0.3015153519871841,
292
+ 0.34730986869917346
293
+ ]
294
+ },
295
+ "gripper_close": {
296
+ "max": [
297
+ 1.0
298
+ ],
299
+ "min": [
300
+ -1.0
301
+ ],
302
+ "mean": [
303
+ -0.33422761717726496
304
+ ],
305
+ "std": [
306
+ 0.9425001619711008
307
+ ],
308
+ "q01": [
309
+ -1.0
310
+ ],
311
+ "q99": [
312
+ 0.681821585371573
313
+ ]
314
+ }
315
+ }
316
+ },
317
+ "modalities": {
318
+ "video": {
319
+ "robot0_eye_in_hand": {
320
+ "resolution": [
321
+ 256,
322
+ 256
323
+ ],
324
+ "channels": 3,
325
+ "fps": 20.0
326
+ },
327
+ "robot0_agentview_left": {
328
+ "resolution": [
329
+ 256,
330
+ 256
331
+ ],
332
+ "channels": 3,
333
+ "fps": 20.0
334
+ },
335
+ "robot0_agentview_right": {
336
+ "resolution": [
337
+ 256,
338
+ 256
339
+ ],
340
+ "channels": 3,
341
+ "fps": 20.0
342
+ }
343
+ },
344
+ "state": {
345
+ "base_position": {
346
+ "absolute": true,
347
+ "rotation_type": null,
348
+ "shape": [
349
+ 3
350
+ ],
351
+ "continuous": true
352
+ },
353
+ "base_rotation": {
354
+ "absolute": true,
355
+ "rotation_type": "quaternion",
356
+ "shape": [
357
+ 4
358
+ ],
359
+ "continuous": true
360
+ },
361
+ "end_effector_position_relative": {
362
+ "absolute": true,
363
+ "rotation_type": null,
364
+ "shape": [
365
+ 3
366
+ ],
367
+ "continuous": true
368
+ },
369
+ "end_effector_rotation_relative": {
370
+ "absolute": true,
371
+ "rotation_type": "quaternion",
372
+ "shape": [
373
+ 4
374
+ ],
375
+ "continuous": true
376
+ },
377
+ "gripper_qpos": {
378
+ "absolute": true,
379
+ "rotation_type": null,
380
+ "shape": [
381
+ 2
382
+ ],
383
+ "continuous": true
384
+ }
385
+ },
386
+ "action": {
387
+ "base_motion": {
388
+ "absolute": true,
389
+ "rotation_type": null,
390
+ "shape": [
391
+ 4
392
+ ],
393
+ "continuous": true
394
+ },
395
+ "control_mode": {
396
+ "absolute": true,
397
+ "rotation_type": null,
398
+ "shape": [
399
+ 1
400
+ ],
401
+ "continuous": true
402
+ },
403
+ "end_effector_position": {
404
+ "absolute": true,
405
+ "rotation_type": null,
406
+ "shape": [
407
+ 3
408
+ ],
409
+ "continuous": true
410
+ },
411
+ "end_effector_rotation": {
412
+ "absolute": true,
413
+ "rotation_type": "axis_angle",
414
+ "shape": [
415
+ 3
416
+ ],
417
+ "continuous": true
418
+ },
419
+ "gripper_close": {
420
+ "absolute": true,
421
+ "rotation_type": null,
422
+ "shape": [
423
+ 1
424
+ ],
425
+ "continuous": true
426
+ }
427
+ }
428
+ },
429
+ "embodiment_tag": "new_embodiment"
430
+ }
431
+ }
RKDv3/base120_p2-40k_p3-80k_fm1-rkd0p5_atomic65_human/default/phase3/model-00001-of-00002.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5be9d3e52b412befbd0e893296605df7c70588d2df8ea3b20fcc419636cdd6f8
3
+ size 4999367032
RKDv3/base120_p2-40k_p3-80k_fm1-rkd0p5_atomic65_human/default/phase3/model-00002-of-00002.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7f1815b389cffe82e2cedf87c0a3450b62f4268306d778cf5d56bdfb3047d287
3
+ size 2664461700
RKDv3/base120_p2-40k_p3-80k_fm1-rkd0p5_atomic65_human/default/phase3/model.safetensors.index.json ADDED
The diff for this file is too large to render. See raw diff
 
RKDv3/base120_p2-40k_p3-80k_fm1-rkd0p5_atomic65_human/default/phase3/trainer_state.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6d874dd47654e6febdf8efa8fb687d576c40884d3ef74704590bdb15b5ea8db5
3
+ size 14938163
RKDv3/base120_p2-40k_p3-80k_fm1-rkd0p5_atomic65_human/default/phase3/training_args.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:cab682ec6f389a5d94312b52adb8f96130271e4abc1e408e26e30808d05a48ac
3
+ size 6033
RKDv3/base120_p2-40k_p3-80k_fm1-rkd0p5_atomic65_human/default/resolved_config.yaml ADDED
@@ -0,0 +1,50 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ experiment_name: RKDv3 TimewarpVAE DA, Student:APHead from GR00T base to 120k on Atomic65
2
+ human, phase2 40k + phase3 80k, FM:RKD 1:0.5
3
+ policy_type: groot_RKDv3
4
+ base_model_path: nvidia/GR00T-N1.5-3B
5
+ output_root: ${HOME}/groot_robocasa/robocasa_v2/RKDv3/base120_p2-40k_p3-80k_fm1-rkd0p5_atomic65_human
6
+ dataset:
7
+ dataset_soup: atomic_pretrain_human
8
+ training:
9
+ num_gpus: 1
10
+ batch_size: 128
11
+ seed: 42
12
+ model:
13
+ rkd_teacher_source: timewarp_action
14
+ rkd_timewarp_vae_checkpoint: ${HOME}/groot_robocasa/TimeWarping/TimewarpVAE/atomic65_z8_g16_h32_tw32_reg0p01_gb_ctrl_30k/checkpoint-step-30000-epoch-2.73
15
+ rkd_relation_mode: flatten
16
+ rkd_loss_type: distance_angle
17
+ rkd_student_reconstruction_head_pooling: attention
18
+ rkd_student_reconstruction_head_enabled: true
19
+ rkd_student_reconstruction_head_dim: 512
20
+ rkd_student_reconstruction_head_hidden_dim: 512
21
+ rkd_distance_loss_weight: 1.0
22
+ rkd_angle_loss_weight: 2.0
23
+ rkd_exclude_diagonal: true
24
+ phases:
25
+ - name: phase2_rkdv3_timewarp_vae_da_aphead_only_40k
26
+ max_steps: 40000
27
+ save_steps: 0
28
+ trainable:
29
+ preset: processing_line_only
30
+ tune_llm: false
31
+ tune_visual: false
32
+ tune_projector: false
33
+ tune_diffusion_model: false
34
+ losses:
35
+ rkd_enabled: true
36
+ rkd_fm_loss_weight: 0.0
37
+ rkd_loss_weight: 1.0
38
+ - name: phase3_fm_rkdv3_timewarp_vae_da_aphead_fixed_0p5_80k
39
+ max_steps: 80000
40
+ save_steps: 0
41
+ trainable:
42
+ tune_llm: false
43
+ tune_visual: false
44
+ tune_projector: true
45
+ tune_diffusion_model: true
46
+ losses:
47
+ rkd_enabled: true
48
+ rkd_fm_loss_weight: 1.0
49
+ rkd_loss_weight: 0.5
50
+ resolved_sweep: {}
RKDv3/base120_p2-60k_p3-60k_fm1-rkd0p5_atomic65_human.stopped_nospace_20260708_204923/default/base120_p2-60k_p3-60k_fm1-rkd0p5_atomic65_human.yaml ADDED
@@ -0,0 +1,49 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ experiment_name: RKDv3 TimewarpVAE DA, Student:APHead from GR00T base to 120k on Atomic65
2
+ human, phase2 60k + phase3 60k, FM:RKD 1:0.5
3
+ policy_type: groot_RKDv3
4
+ base_model_path: nvidia/GR00T-N1.5-3B
5
+ output_root: ${HOME}/groot_robocasa/robocasa_v2/RKDv3/base120_p2-60k_p3-60k_fm1-rkd0p5_atomic65_human
6
+ dataset:
7
+ dataset_soup: atomic_pretrain_human
8
+ training:
9
+ num_gpus: 1
10
+ batch_size: 128
11
+ seed: 42
12
+ model:
13
+ rkd_teacher_source: timewarp_action
14
+ rkd_timewarp_vae_checkpoint: ${HOME}/groot_robocasa/TimeWarping/TimewarpVAE/robocasa_vanilla_t16_z8_g8_h32_reg0p01_beta0p01/checkpoint-step-60000-epoch-17.46
15
+ rkd_relation_mode: flatten
16
+ rkd_loss_type: distance_angle
17
+ rkd_student_reconstruction_head_pooling: attention
18
+ rkd_student_reconstruction_head_enabled: true
19
+ rkd_student_reconstruction_head_dim: 512
20
+ rkd_student_reconstruction_head_hidden_dim: 512
21
+ rkd_distance_loss_weight: 1.0
22
+ rkd_angle_loss_weight: 2.0
23
+ rkd_exclude_diagonal: true
24
+ phases:
25
+ - name: phase2_rkdv3_timewarp_vae_da_aphead_only_60k
26
+ max_steps: 60000
27
+ save_steps: 0
28
+ trainable:
29
+ preset: processing_line_only
30
+ tune_llm: false
31
+ tune_visual: false
32
+ tune_projector: false
33
+ tune_diffusion_model: false
34
+ losses:
35
+ rkd_enabled: true
36
+ rkd_fm_loss_weight: 0.0
37
+ rkd_loss_weight: 1.0
38
+ - name: phase3_fm_rkdv3_timewarp_vae_da_aphead_fixed_0p5_60k
39
+ max_steps: 60000
40
+ save_steps: 0
41
+ trainable:
42
+ tune_llm: false
43
+ tune_visual: false
44
+ tune_projector: true
45
+ tune_diffusion_model: true
46
+ losses:
47
+ rkd_enabled: true
48
+ rkd_fm_loss_weight: 1.0
49
+ rkd_loss_weight: 0.5
RKDv3/base120_p2-60k_p3-60k_fm1-rkd0p5_atomic65_human.stopped_nospace_20260708_204923/default/phase2/experiment_cfg/metadata.json ADDED
@@ -0,0 +1,431 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "new_embodiment": {
3
+ "statistics": {
4
+ "state": {
5
+ "base_position": {
6
+ "max": [
7
+ 7.778976917266846,
8
+ 0.4876587688922882,
9
+ 0.7223060727119446
10
+ ],
11
+ "min": [
12
+ -4.94293737411499,
13
+ -6.890198230743408,
14
+ 0.6994727253913879
15
+ ],
16
+ "mean": [
17
+ 2.645730005800653,
18
+ -2.110066447108802,
19
+ 0.7004254938507055
20
+ ],
21
+ "std": [
22
+ 1.674798582069234,
23
+ 1.350392482318102,
24
+ 0.001107956680671985
25
+ ],
26
+ "q01": [
27
+ -1.0132281893317567,
28
+ -5.366714195503661,
29
+ 0.699973412653353
30
+ ],
31
+ "q99": [
32
+ 6.099414437993715,
33
+ -0.5498886012762193,
34
+ 0.7032860957123939
35
+ ]
36
+ },
37
+ "base_rotation": {
38
+ "max": [
39
+ 0.0,
40
+ 0.0,
41
+ 1.0,
42
+ 1.0
43
+ ],
44
+ "min": [
45
+ 0.0,
46
+ 0.0,
47
+ -1.0,
48
+ 0.0
49
+ ],
50
+ "mean": [
51
+ 0.0,
52
+ 0.0,
53
+ 0.24103797327895668,
54
+ 0.6071393324239354
55
+ ],
56
+ "std": [
57
+ 0.0,
58
+ 0.0,
59
+ 0.6615962784298529,
60
+ 0.36814045893518177
61
+ ],
62
+ "q01": [
63
+ 0.0,
64
+ 0.0,
65
+ -1.0,
66
+ 3.3913522062507517e-06
67
+ ],
68
+ "q99": [
69
+ 0.0,
70
+ 0.0,
71
+ 1.0,
72
+ 1.0
73
+ ]
74
+ },
75
+ "end_effector_position_relative": {
76
+ "max": [
77
+ 0.9014528393745422,
78
+ 0.8569459915161133,
79
+ 1.053884744644165
80
+ ],
81
+ "min": [
82
+ -0.5828549861907959,
83
+ -0.8486236333847046,
84
+ -0.3446122705936432
85
+ ],
86
+ "mean": [
87
+ 0.2641355221895152,
88
+ -0.01768191244102862,
89
+ 0.4458847037294584
90
+ ],
91
+ "std": [
92
+ 0.16437879912713016,
93
+ 0.24835923712456218,
94
+ 0.2124190803465676
95
+ ],
96
+ "q01": [
97
+ -0.004962867024380738,
98
+ -0.46862868595225443,
99
+ 0.20316236863116172
100
+ ],
101
+ "q99": [
102
+ 0.5495510296644974,
103
+ 0.44201455499146414,
104
+ 0.6984316542098391
105
+ ]
106
+ },
107
+ "end_effector_rotation_relative": {
108
+ "max": [
109
+ 0.9999998807907104,
110
+ 0.9997354745864868,
111
+ 0.9900747537612915,
112
+ 0.9735639095306396
113
+ ],
114
+ "min": [
115
+ -0.9999930262565613,
116
+ -0.9999884366989136,
117
+ -0.9975369572639465,
118
+ 6.754255821306288e-08
119
+ ],
120
+ "mean": [
121
+ -0.23536101367097556,
122
+ 0.0384247508727208,
123
+ -0.06854891687606551,
124
+ 0.1400254132624415
125
+ ],
126
+ "std": [
127
+ 0.8296216819146821,
128
+ 0.30067208350163366,
129
+ 0.33390768096919515,
130
+ 0.16926885514344236
131
+ ],
132
+ "q01": [
133
+ -0.9950540525525874,
134
+ -0.5590028004550295,
135
+ -0.5635208777456633,
136
+ 0.001747048046951923
137
+ ],
138
+ "q99": [
139
+ 0.9472459325196646,
140
+ 0.6183763159959191,
141
+ 0.5108060516185887,
142
+ 0.46646525101830405
143
+ ]
144
+ },
145
+ "gripper_qpos": {
146
+ "max": [
147
+ 0.055869169533252716,
148
+ 0.010916369967162609
149
+ ],
150
+ "min": [
151
+ -0.011669117026031017,
152
+ -0.06307756155729294
153
+ ],
154
+ "mean": [
155
+ 0.031849369408926076,
156
+ -0.03200748595341724
157
+ ],
158
+ "std": [
159
+ 0.012248040928874927,
160
+ 0.01206276030599065
161
+ ],
162
+ "q01": [
163
+ 0.006329614505051435,
164
+ -0.0405872199089872
165
+ ],
166
+ "q99": [
167
+ 0.040445365520021676,
168
+ -0.006282430560590391
169
+ ]
170
+ }
171
+ },
172
+ "action": {
173
+ "base_motion": {
174
+ "max": [
175
+ 1.0,
176
+ 1.0,
177
+ 1.0,
178
+ 0.0
179
+ ],
180
+ "min": [
181
+ -1.0,
182
+ -1.0,
183
+ -1.0,
184
+ 0.0
185
+ ],
186
+ "mean": [
187
+ 0.0071452133647496785,
188
+ 0.0015603604277540566,
189
+ -0.0014052511062474704,
190
+ 0.0
191
+ ],
192
+ "std": [
193
+ 0.11471999232420287,
194
+ 0.10209228015027566,
195
+ 0.0855718789814838,
196
+ 0.0
197
+ ],
198
+ "q01": [
199
+ -0.13329522793794643,
200
+ -0.16270717101832055,
201
+ -0.12619798329427676,
202
+ 0.0
203
+ ],
204
+ "q99": [
205
+ 0.1721366453818688,
206
+ 0.1972873452547982,
207
+ 0.11128571706089413,
208
+ 0.0
209
+ ]
210
+ },
211
+ "control_mode": {
212
+ "max": [
213
+ 1.0
214
+ ],
215
+ "min": [
216
+ -1.0
217
+ ],
218
+ "mean": [
219
+ -0.9218734936549906
220
+ ],
221
+ "std": [
222
+ 0.3874975703570591
223
+ ],
224
+ "q01": [
225
+ -1.0
226
+ ],
227
+ "q99": [
228
+ -0.26692269247002653
229
+ ]
230
+ },
231
+ "end_effector_position": {
232
+ "max": [
233
+ 1.0,
234
+ 1.0,
235
+ 1.0
236
+ ],
237
+ "min": [
238
+ -1.0,
239
+ -1.0,
240
+ -1.0
241
+ ],
242
+ "mean": [
243
+ 0.002920238070268553,
244
+ -0.0060913649045883315,
245
+ -0.07322447393298565
246
+ ],
247
+ "std": [
248
+ 0.4482114094812734,
249
+ 0.41233706159751987,
250
+ 0.4006084454909327
251
+ ],
252
+ "q01": [
253
+ -0.934104737554819,
254
+ -0.9255792016234818,
255
+ -0.8085822192859831
256
+ ],
257
+ "q99": [
258
+ 0.7874185230421324,
259
+ 0.9004049161564378,
260
+ 0.8018332495119154
261
+ ]
262
+ },
263
+ "end_effector_rotation": {
264
+ "max": [
265
+ 1.0,
266
+ 1.0,
267
+ 1.0
268
+ ],
269
+ "min": [
270
+ -1.0,
271
+ -1.0,
272
+ -1.0
273
+ ],
274
+ "mean": [
275
+ 0.0077772595328458915,
276
+ -0.022620543966259083,
277
+ -0.00029332176634943376
278
+ ],
279
+ "std": [
280
+ 0.11155576533045976,
281
+ 0.13202104364511658,
282
+ 0.1240098754411023
283
+ ],
284
+ "q01": [
285
+ -0.2726786325766535,
286
+ -0.4125887853047325,
287
+ -0.3445183934868137
288
+ ],
289
+ "q99": [
290
+ 0.31803481867756295,
291
+ 0.3015153519871841,
292
+ 0.34730986869917346
293
+ ]
294
+ },
295
+ "gripper_close": {
296
+ "max": [
297
+ 1.0
298
+ ],
299
+ "min": [
300
+ -1.0
301
+ ],
302
+ "mean": [
303
+ -0.33422761717726496
304
+ ],
305
+ "std": [
306
+ 0.9425001619711008
307
+ ],
308
+ "q01": [
309
+ -1.0
310
+ ],
311
+ "q99": [
312
+ 0.681821585371573
313
+ ]
314
+ }
315
+ }
316
+ },
317
+ "modalities": {
318
+ "video": {
319
+ "robot0_eye_in_hand": {
320
+ "resolution": [
321
+ 256,
322
+ 256
323
+ ],
324
+ "channels": 3,
325
+ "fps": 20.0
326
+ },
327
+ "robot0_agentview_left": {
328
+ "resolution": [
329
+ 256,
330
+ 256
331
+ ],
332
+ "channels": 3,
333
+ "fps": 20.0
334
+ },
335
+ "robot0_agentview_right": {
336
+ "resolution": [
337
+ 256,
338
+ 256
339
+ ],
340
+ "channels": 3,
341
+ "fps": 20.0
342
+ }
343
+ },
344
+ "state": {
345
+ "base_position": {
346
+ "absolute": true,
347
+ "rotation_type": null,
348
+ "shape": [
349
+ 3
350
+ ],
351
+ "continuous": true
352
+ },
353
+ "base_rotation": {
354
+ "absolute": true,
355
+ "rotation_type": "quaternion",
356
+ "shape": [
357
+ 4
358
+ ],
359
+ "continuous": true
360
+ },
361
+ "end_effector_position_relative": {
362
+ "absolute": true,
363
+ "rotation_type": null,
364
+ "shape": [
365
+ 3
366
+ ],
367
+ "continuous": true
368
+ },
369
+ "end_effector_rotation_relative": {
370
+ "absolute": true,
371
+ "rotation_type": "quaternion",
372
+ "shape": [
373
+ 4
374
+ ],
375
+ "continuous": true
376
+ },
377
+ "gripper_qpos": {
378
+ "absolute": true,
379
+ "rotation_type": null,
380
+ "shape": [
381
+ 2
382
+ ],
383
+ "continuous": true
384
+ }
385
+ },
386
+ "action": {
387
+ "base_motion": {
388
+ "absolute": true,
389
+ "rotation_type": null,
390
+ "shape": [
391
+ 4
392
+ ],
393
+ "continuous": true
394
+ },
395
+ "control_mode": {
396
+ "absolute": true,
397
+ "rotation_type": null,
398
+ "shape": [
399
+ 1
400
+ ],
401
+ "continuous": true
402
+ },
403
+ "end_effector_position": {
404
+ "absolute": true,
405
+ "rotation_type": null,
406
+ "shape": [
407
+ 3
408
+ ],
409
+ "continuous": true
410
+ },
411
+ "end_effector_rotation": {
412
+ "absolute": true,
413
+ "rotation_type": "axis_angle",
414
+ "shape": [
415
+ 3
416
+ ],
417
+ "continuous": true
418
+ },
419
+ "gripper_close": {
420
+ "absolute": true,
421
+ "rotation_type": null,
422
+ "shape": [
423
+ 1
424
+ ],
425
+ "continuous": true
426
+ }
427
+ }
428
+ },
429
+ "embodiment_tag": "new_embodiment"
430
+ }
431
+ }
RKDv3/base120_p2-60k_p3-60k_fm1-rkd0p5_atomic65_human.stopped_nospace_20260708_204923/default/resolved_config.yaml ADDED
@@ -0,0 +1,50 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ experiment_name: RKDv3 TimewarpVAE DA, Student:APHead from GR00T base to 120k on Atomic65
2
+ human, phase2 60k + phase3 60k, FM:RKD 1:0.5
3
+ policy_type: groot_RKDv3
4
+ base_model_path: nvidia/GR00T-N1.5-3B
5
+ output_root: ${HOME}/groot_robocasa/robocasa_v2/RKDv3/base120_p2-60k_p3-60k_fm1-rkd0p5_atomic65_human
6
+ dataset:
7
+ dataset_soup: atomic_pretrain_human
8
+ training:
9
+ num_gpus: 1
10
+ batch_size: 128
11
+ seed: 42
12
+ model:
13
+ rkd_teacher_source: timewarp_action
14
+ rkd_timewarp_vae_checkpoint: ${HOME}/groot_robocasa/TimeWarping/TimewarpVAE/robocasa_vanilla_t16_z8_g8_h32_reg0p01_beta0p01/checkpoint-step-60000-epoch-17.46
15
+ rkd_relation_mode: flatten
16
+ rkd_loss_type: distance_angle
17
+ rkd_student_reconstruction_head_pooling: attention
18
+ rkd_student_reconstruction_head_enabled: true
19
+ rkd_student_reconstruction_head_dim: 512
20
+ rkd_student_reconstruction_head_hidden_dim: 512
21
+ rkd_distance_loss_weight: 1.0
22
+ rkd_angle_loss_weight: 2.0
23
+ rkd_exclude_diagonal: true
24
+ phases:
25
+ - name: phase2_rkdv3_timewarp_vae_da_aphead_only_60k
26
+ max_steps: 60000
27
+ save_steps: 0
28
+ trainable:
29
+ preset: processing_line_only
30
+ tune_llm: false
31
+ tune_visual: false
32
+ tune_projector: false
33
+ tune_diffusion_model: false
34
+ losses:
35
+ rkd_enabled: true
36
+ rkd_fm_loss_weight: 0.0
37
+ rkd_loss_weight: 1.0
38
+ - name: phase3_fm_rkdv3_timewarp_vae_da_aphead_fixed_0p5_60k
39
+ max_steps: 60000
40
+ save_steps: 0
41
+ trainable:
42
+ tune_llm: false
43
+ tune_visual: false
44
+ tune_projector: true
45
+ tune_diffusion_model: true
46
+ losses:
47
+ rkd_enabled: true
48
+ rkd_fm_loss_weight: 1.0
49
+ rkd_loss_weight: 0.5
50
+ resolved_sweep: {}
RKDv3/base120_p2-60k_p3-60k_fm1-rkd0p5_atomic65_human/default/base120_p2-60k_p3-60k_fm1-rkd0p5_atomic65_human.yaml ADDED
@@ -0,0 +1,49 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ experiment_name: RKDv3 TimewarpVAE DA, Student:APHead from GR00T base to 120k on Atomic65
2
+ human, phase2 60k + phase3 60k, FM:RKD 1:0.5
3
+ policy_type: groot_RKDv3
4
+ base_model_path: nvidia/GR00T-N1.5-3B
5
+ output_root: ${HOME}/groot_robocasa/robocasa_v2/RKDv3/base120_p2-60k_p3-60k_fm1-rkd0p5_atomic65_human
6
+ dataset:
7
+ dataset_soup: atomic_pretrain_human
8
+ training:
9
+ num_gpus: 1
10
+ batch_size: 128
11
+ seed: 42
12
+ model:
13
+ rkd_teacher_source: timewarp_action
14
+ rkd_timewarp_vae_checkpoint: ${HOME}/groot_robocasa/TimeWarping/TimewarpVAE/atomic65_z8_g16_h32_tw32_reg0p01_gb_ctrl_30k/checkpoint-step-30000-epoch-2.73
15
+ rkd_relation_mode: flatten
16
+ rkd_loss_type: distance_angle
17
+ rkd_student_reconstruction_head_pooling: attention
18
+ rkd_student_reconstruction_head_enabled: true
19
+ rkd_student_reconstruction_head_dim: 512
20
+ rkd_student_reconstruction_head_hidden_dim: 512
21
+ rkd_distance_loss_weight: 1.0
22
+ rkd_angle_loss_weight: 2.0
23
+ rkd_exclude_diagonal: true
24
+ phases:
25
+ - name: phase2_rkdv3_timewarp_vae_da_aphead_only_60k
26
+ max_steps: 60000
27
+ save_steps: 0
28
+ trainable:
29
+ preset: processing_line_only
30
+ tune_llm: false
31
+ tune_visual: false
32
+ tune_projector: false
33
+ tune_diffusion_model: false
34
+ losses:
35
+ rkd_enabled: true
36
+ rkd_fm_loss_weight: 0.0
37
+ rkd_loss_weight: 1.0
38
+ - name: phase3_fm_rkdv3_timewarp_vae_da_aphead_fixed_0p5_60k
39
+ max_steps: 60000
40
+ save_steps: 0
41
+ trainable:
42
+ tune_llm: false
43
+ tune_visual: false
44
+ tune_projector: true
45
+ tune_diffusion_model: true
46
+ losses:
47
+ rkd_enabled: true
48
+ rkd_fm_loss_weight: 1.0
49
+ rkd_loss_weight: 0.5
RKDv3/base120_p2-60k_p3-60k_fm1-rkd0p5_atomic65_human/default/phase2/config.json ADDED
@@ -0,0 +1,101 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "action_dim": 32,
3
+ "action_head_cfg": {
4
+ "action_dim": 32,
5
+ "action_horizon": 16,
6
+ "add_pos_embed": true,
7
+ "backbone_embedding_dim": 2048,
8
+ "diffusion_model_cfg": {
9
+ "attention_head_dim": 48,
10
+ "cross_attention_dim": 2048,
11
+ "dropout": 0.2,
12
+ "final_dropout": true,
13
+ "interleave_self_attention": true,
14
+ "norm_type": "ada_norm",
15
+ "num_attention_heads": 32,
16
+ "num_layers": 16,
17
+ "output_dim": 1024,
18
+ "positional_embeddings": null
19
+ },
20
+ "hidden_size": 1024,
21
+ "input_embedding_dim": 1536,
22
+ "max_action_dim": 32,
23
+ "max_state_dim": 64,
24
+ "model_dtype": "float32",
25
+ "noise_beta_alpha": 1.5,
26
+ "noise_beta_beta": 1.0,
27
+ "noise_s": 0.999,
28
+ "num_inference_timesteps": 4,
29
+ "num_target_vision_tokens": 32,
30
+ "num_timestep_buckets": 1000,
31
+ "tune_diffusion_model": true,
32
+ "tune_projector": true,
33
+ "use_vlln": true,
34
+ "vl_self_attention_cfg": {
35
+ "attention_head_dim": 64,
36
+ "dropout": 0.2,
37
+ "final_dropout": true,
38
+ "num_attention_heads": 32,
39
+ "num_layers": 4,
40
+ "positional_embeddings": null
41
+ }
42
+ },
43
+ "action_horizon": 16,
44
+ "architectures": [
45
+ "GR00T_N1_5_RKD"
46
+ ],
47
+ "attn_implementation": null,
48
+ "backbone_cfg": {
49
+ "eagle_path": "NVEagle/eagle_er-qwen3_1_7B-Siglip2_400M_stage1_5_128gpu_er_v7_1mlp_nops",
50
+ "load_bf16": false,
51
+ "project_to_dim": null,
52
+ "reproject_vision": false,
53
+ "select_layer": 12,
54
+ "tune_llm": false,
55
+ "tune_visual": true,
56
+ "use_flash_attention": true
57
+ },
58
+ "compute_dtype": "bfloat16",
59
+ "hidden_size": 2048,
60
+ "model_dtype": "float32",
61
+ "model_type": "gr00t_n1_5",
62
+ "rkd_action_encoder_cls_ffn_hidden_dim": 2048,
63
+ "rkd_action_encoder_cls_num_heads": 8,
64
+ "rkd_action_encoder_cls_num_layers": 1,
65
+ "rkd_action_encoder_projector_dim": 512,
66
+ "rkd_action_encoder_projector_enabled": false,
67
+ "rkd_action_encoder_projector_pooling": "flatten",
68
+ "rkd_action_temp": 0.1,
69
+ "rkd_angle_loss_weight": 2.0,
70
+ "rkd_distance_loss_weight": 1.0,
71
+ "rkd_dtw_adaptive_floor": 0.05,
72
+ "rkd_dtw_adaptive_mix_lambda": 0.5,
73
+ "rkd_dtw_base_weight": 1.0,
74
+ "rkd_dtw_gripper_weight": 0.5,
75
+ "rkd_dtw_mix_mode": "distance_mix",
76
+ "rkd_dtw_normalize_components": true,
77
+ "rkd_dtw_pos_weight": 1.0,
78
+ "rkd_dtw_rot_weight": 1.0,
79
+ "rkd_dtw_timing_log_interval": 1000,
80
+ "rkd_enabled": true,
81
+ "rkd_exclude_diagonal": true,
82
+ "rkd_fm_loss_weight": 0.0,
83
+ "rkd_loss_type": "distance_angle",
84
+ "rkd_loss_weight": 1.0,
85
+ "rkd_loss_weight_end": 0.0,
86
+ "rkd_loss_weight_schedule": null,
87
+ "rkd_loss_weight_start": 0.1,
88
+ "rkd_raw_action_projector_dim": 512,
89
+ "rkd_raw_action_projector_enabled": false,
90
+ "rkd_raw_action_projector_pooling": "flatten",
91
+ "rkd_relation_mode": "flatten",
92
+ "rkd_student_reconstruction_head_dim": 512,
93
+ "rkd_student_reconstruction_head_enabled": true,
94
+ "rkd_student_reconstruction_head_hidden_dim": 512,
95
+ "rkd_student_reconstruction_head_pooling": "attention",
96
+ "rkd_teacher_source": "timewarp_action",
97
+ "rkd_timewarp_vae_checkpoint": "${HOME}/groot_robocasa/TimeWarping/TimewarpVAE/atomic65_z8_g16_h32_tw32_reg0p01_gb_ctrl_30k/checkpoint-step-30000-epoch-2.73",
98
+ "rkd_vlm_temp": 0.1,
99
+ "torch_dtype": "bfloat16",
100
+ "transformers_version": "4.51.3"
101
+ }
RKDv3/base120_p2-60k_p3-60k_fm1-rkd0p5_atomic65_human/default/phase2/experiment_cfg/metadata.json ADDED
@@ -0,0 +1,431 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "new_embodiment": {
3
+ "statistics": {
4
+ "state": {
5
+ "base_position": {
6
+ "max": [
7
+ 7.778976917266846,
8
+ 0.4876587688922882,
9
+ 0.7223060727119446
10
+ ],
11
+ "min": [
12
+ -4.94293737411499,
13
+ -6.890198230743408,
14
+ 0.6994727253913879
15
+ ],
16
+ "mean": [
17
+ 2.645730005800653,
18
+ -2.110066447108802,
19
+ 0.7004254938507055
20
+ ],
21
+ "std": [
22
+ 1.674798582069234,
23
+ 1.350392482318102,
24
+ 0.001107956680671985
25
+ ],
26
+ "q01": [
27
+ -1.0132281893317567,
28
+ -5.366714195503661,
29
+ 0.699973412653353
30
+ ],
31
+ "q99": [
32
+ 6.099414437993715,
33
+ -0.5498886012762193,
34
+ 0.7032860957123939
35
+ ]
36
+ },
37
+ "base_rotation": {
38
+ "max": [
39
+ 0.0,
40
+ 0.0,
41
+ 1.0,
42
+ 1.0
43
+ ],
44
+ "min": [
45
+ 0.0,
46
+ 0.0,
47
+ -1.0,
48
+ 0.0
49
+ ],
50
+ "mean": [
51
+ 0.0,
52
+ 0.0,
53
+ 0.24103797327895668,
54
+ 0.6071393324239354
55
+ ],
56
+ "std": [
57
+ 0.0,
58
+ 0.0,
59
+ 0.6615962784298529,
60
+ 0.36814045893518177
61
+ ],
62
+ "q01": [
63
+ 0.0,
64
+ 0.0,
65
+ -1.0,
66
+ 3.3913522062507517e-06
67
+ ],
68
+ "q99": [
69
+ 0.0,
70
+ 0.0,
71
+ 1.0,
72
+ 1.0
73
+ ]
74
+ },
75
+ "end_effector_position_relative": {
76
+ "max": [
77
+ 0.9014528393745422,
78
+ 0.8569459915161133,
79
+ 1.053884744644165
80
+ ],
81
+ "min": [
82
+ -0.5828549861907959,
83
+ -0.8486236333847046,
84
+ -0.3446122705936432
85
+ ],
86
+ "mean": [
87
+ 0.2641355221895152,
88
+ -0.01768191244102862,
89
+ 0.4458847037294584
90
+ ],
91
+ "std": [
92
+ 0.16437879912713016,
93
+ 0.24835923712456218,
94
+ 0.2124190803465676
95
+ ],
96
+ "q01": [
97
+ -0.004962867024380738,
98
+ -0.46862868595225443,
99
+ 0.20316236863116172
100
+ ],
101
+ "q99": [
102
+ 0.5495510296644974,
103
+ 0.44201455499146414,
104
+ 0.6984316542098391
105
+ ]
106
+ },
107
+ "end_effector_rotation_relative": {
108
+ "max": [
109
+ 0.9999998807907104,
110
+ 0.9997354745864868,
111
+ 0.9900747537612915,
112
+ 0.9735639095306396
113
+ ],
114
+ "min": [
115
+ -0.9999930262565613,
116
+ -0.9999884366989136,
117
+ -0.9975369572639465,
118
+ 6.754255821306288e-08
119
+ ],
120
+ "mean": [
121
+ -0.23536101367097556,
122
+ 0.0384247508727208,
123
+ -0.06854891687606551,
124
+ 0.1400254132624415
125
+ ],
126
+ "std": [
127
+ 0.8296216819146821,
128
+ 0.30067208350163366,
129
+ 0.33390768096919515,
130
+ 0.16926885514344236
131
+ ],
132
+ "q01": [
133
+ -0.9950540525525874,
134
+ -0.5590028004550295,
135
+ -0.5635208777456633,
136
+ 0.001747048046951923
137
+ ],
138
+ "q99": [
139
+ 0.9472459325196646,
140
+ 0.6183763159959191,
141
+ 0.5108060516185887,
142
+ 0.46646525101830405
143
+ ]
144
+ },
145
+ "gripper_qpos": {
146
+ "max": [
147
+ 0.055869169533252716,
148
+ 0.010916369967162609
149
+ ],
150
+ "min": [
151
+ -0.011669117026031017,
152
+ -0.06307756155729294
153
+ ],
154
+ "mean": [
155
+ 0.031849369408926076,
156
+ -0.03200748595341724
157
+ ],
158
+ "std": [
159
+ 0.012248040928874927,
160
+ 0.01206276030599065
161
+ ],
162
+ "q01": [
163
+ 0.006329614505051435,
164
+ -0.0405872199089872
165
+ ],
166
+ "q99": [
167
+ 0.040445365520021676,
168
+ -0.006282430560590391
169
+ ]
170
+ }
171
+ },
172
+ "action": {
173
+ "base_motion": {
174
+ "max": [
175
+ 1.0,
176
+ 1.0,
177
+ 1.0,
178
+ 0.0
179
+ ],
180
+ "min": [
181
+ -1.0,
182
+ -1.0,
183
+ -1.0,
184
+ 0.0
185
+ ],
186
+ "mean": [
187
+ 0.0071452133647496785,
188
+ 0.0015603604277540566,
189
+ -0.0014052511062474704,
190
+ 0.0
191
+ ],
192
+ "std": [
193
+ 0.11471999232420287,
194
+ 0.10209228015027566,
195
+ 0.0855718789814838,
196
+ 0.0
197
+ ],
198
+ "q01": [
199
+ -0.13329522793794643,
200
+ -0.16270717101832055,
201
+ -0.12619798329427676,
202
+ 0.0
203
+ ],
204
+ "q99": [
205
+ 0.1721366453818688,
206
+ 0.1972873452547982,
207
+ 0.11128571706089413,
208
+ 0.0
209
+ ]
210
+ },
211
+ "control_mode": {
212
+ "max": [
213
+ 1.0
214
+ ],
215
+ "min": [
216
+ -1.0
217
+ ],
218
+ "mean": [
219
+ -0.9218734936549906
220
+ ],
221
+ "std": [
222
+ 0.3874975703570591
223
+ ],
224
+ "q01": [
225
+ -1.0
226
+ ],
227
+ "q99": [
228
+ -0.26692269247002653
229
+ ]
230
+ },
231
+ "end_effector_position": {
232
+ "max": [
233
+ 1.0,
234
+ 1.0,
235
+ 1.0
236
+ ],
237
+ "min": [
238
+ -1.0,
239
+ -1.0,
240
+ -1.0
241
+ ],
242
+ "mean": [
243
+ 0.002920238070268553,
244
+ -0.0060913649045883315,
245
+ -0.07322447393298565
246
+ ],
247
+ "std": [
248
+ 0.4482114094812734,
249
+ 0.41233706159751987,
250
+ 0.4006084454909327
251
+ ],
252
+ "q01": [
253
+ -0.934104737554819,
254
+ -0.9255792016234818,
255
+ -0.8085822192859831
256
+ ],
257
+ "q99": [
258
+ 0.7874185230421324,
259
+ 0.9004049161564378,
260
+ 0.8018332495119154
261
+ ]
262
+ },
263
+ "end_effector_rotation": {
264
+ "max": [
265
+ 1.0,
266
+ 1.0,
267
+ 1.0
268
+ ],
269
+ "min": [
270
+ -1.0,
271
+ -1.0,
272
+ -1.0
273
+ ],
274
+ "mean": [
275
+ 0.0077772595328458915,
276
+ -0.022620543966259083,
277
+ -0.00029332176634943376
278
+ ],
279
+ "std": [
280
+ 0.11155576533045976,
281
+ 0.13202104364511658,
282
+ 0.1240098754411023
283
+ ],
284
+ "q01": [
285
+ -0.2726786325766535,
286
+ -0.4125887853047325,
287
+ -0.3445183934868137
288
+ ],
289
+ "q99": [
290
+ 0.31803481867756295,
291
+ 0.3015153519871841,
292
+ 0.34730986869917346
293
+ ]
294
+ },
295
+ "gripper_close": {
296
+ "max": [
297
+ 1.0
298
+ ],
299
+ "min": [
300
+ -1.0
301
+ ],
302
+ "mean": [
303
+ -0.33422761717726496
304
+ ],
305
+ "std": [
306
+ 0.9425001619711008
307
+ ],
308
+ "q01": [
309
+ -1.0
310
+ ],
311
+ "q99": [
312
+ 0.681821585371573
313
+ ]
314
+ }
315
+ }
316
+ },
317
+ "modalities": {
318
+ "video": {
319
+ "robot0_eye_in_hand": {
320
+ "resolution": [
321
+ 256,
322
+ 256
323
+ ],
324
+ "channels": 3,
325
+ "fps": 20.0
326
+ },
327
+ "robot0_agentview_left": {
328
+ "resolution": [
329
+ 256,
330
+ 256
331
+ ],
332
+ "channels": 3,
333
+ "fps": 20.0
334
+ },
335
+ "robot0_agentview_right": {
336
+ "resolution": [
337
+ 256,
338
+ 256
339
+ ],
340
+ "channels": 3,
341
+ "fps": 20.0
342
+ }
343
+ },
344
+ "state": {
345
+ "base_position": {
346
+ "absolute": true,
347
+ "rotation_type": null,
348
+ "shape": [
349
+ 3
350
+ ],
351
+ "continuous": true
352
+ },
353
+ "base_rotation": {
354
+ "absolute": true,
355
+ "rotation_type": "quaternion",
356
+ "shape": [
357
+ 4
358
+ ],
359
+ "continuous": true
360
+ },
361
+ "end_effector_position_relative": {
362
+ "absolute": true,
363
+ "rotation_type": null,
364
+ "shape": [
365
+ 3
366
+ ],
367
+ "continuous": true
368
+ },
369
+ "end_effector_rotation_relative": {
370
+ "absolute": true,
371
+ "rotation_type": "quaternion",
372
+ "shape": [
373
+ 4
374
+ ],
375
+ "continuous": true
376
+ },
377
+ "gripper_qpos": {
378
+ "absolute": true,
379
+ "rotation_type": null,
380
+ "shape": [
381
+ 2
382
+ ],
383
+ "continuous": true
384
+ }
385
+ },
386
+ "action": {
387
+ "base_motion": {
388
+ "absolute": true,
389
+ "rotation_type": null,
390
+ "shape": [
391
+ 4
392
+ ],
393
+ "continuous": true
394
+ },
395
+ "control_mode": {
396
+ "absolute": true,
397
+ "rotation_type": null,
398
+ "shape": [
399
+ 1
400
+ ],
401
+ "continuous": true
402
+ },
403
+ "end_effector_position": {
404
+ "absolute": true,
405
+ "rotation_type": null,
406
+ "shape": [
407
+ 3
408
+ ],
409
+ "continuous": true
410
+ },
411
+ "end_effector_rotation": {
412
+ "absolute": true,
413
+ "rotation_type": "axis_angle",
414
+ "shape": [
415
+ 3
416
+ ],
417
+ "continuous": true
418
+ },
419
+ "gripper_close": {
420
+ "absolute": true,
421
+ "rotation_type": null,
422
+ "shape": [
423
+ 1
424
+ ],
425
+ "continuous": true
426
+ }
427
+ }
428
+ },
429
+ "embodiment_tag": "new_embodiment"
430
+ }
431
+ }
RKDv3/base120_p2-60k_p3-60k_fm1-rkd0p5_atomic65_human/default/phase2/model-00001-of-00002.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0e65895a7b6c052e58b679bdd68bff54a68694a81f38e23d3e03a92059b3600b
3
+ size 4999367032
RKDv3/base120_p2-60k_p3-60k_fm1-rkd0p5_atomic65_human/default/phase2/model-00002-of-00002.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1f180b376369d5c5550806f3c58db54714fd0f986cd41593a589ba29ec26e3fe
3
+ size 2664461700
RKDv3/base120_p2-60k_p3-60k_fm1-rkd0p5_atomic65_human/default/phase2/model.safetensors.index.json ADDED
The diff for this file is too large to render. See raw diff
 
RKDv3/base120_p2-60k_p3-60k_fm1-rkd0p5_atomic65_human/default/phase2/trainer_state.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a93dd5c979d7e9de8a311f6ddd355187eb19314049c8773731f6dcaf887841d7
3
+ size 10995731
RKDv3/base120_p2-60k_p3-60k_fm1-rkd0p5_atomic65_human/default/phase2/training_args.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:32020ad7f082328a011d7f08f5d5f138b36d34d35d9d325b589526ba57831432
3
+ size 6033
RKDv3/base120_p2-60k_p3-60k_fm1-rkd0p5_atomic65_human/default/phase3/config.json ADDED
@@ -0,0 +1,101 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "action_dim": 32,
3
+ "action_head_cfg": {
4
+ "action_dim": 32,
5
+ "action_horizon": 16,
6
+ "add_pos_embed": true,
7
+ "backbone_embedding_dim": 2048,
8
+ "diffusion_model_cfg": {
9
+ "attention_head_dim": 48,
10
+ "cross_attention_dim": 2048,
11
+ "dropout": 0.2,
12
+ "final_dropout": true,
13
+ "interleave_self_attention": true,
14
+ "norm_type": "ada_norm",
15
+ "num_attention_heads": 32,
16
+ "num_layers": 16,
17
+ "output_dim": 1024,
18
+ "positional_embeddings": null
19
+ },
20
+ "hidden_size": 1024,
21
+ "input_embedding_dim": 1536,
22
+ "max_action_dim": 32,
23
+ "max_state_dim": 64,
24
+ "model_dtype": "float32",
25
+ "noise_beta_alpha": 1.5,
26
+ "noise_beta_beta": 1.0,
27
+ "noise_s": 0.999,
28
+ "num_inference_timesteps": 4,
29
+ "num_target_vision_tokens": 32,
30
+ "num_timestep_buckets": 1000,
31
+ "tune_diffusion_model": true,
32
+ "tune_projector": true,
33
+ "use_vlln": true,
34
+ "vl_self_attention_cfg": {
35
+ "attention_head_dim": 64,
36
+ "dropout": 0.2,
37
+ "final_dropout": true,
38
+ "num_attention_heads": 32,
39
+ "num_layers": 4,
40
+ "positional_embeddings": null
41
+ }
42
+ },
43
+ "action_horizon": 16,
44
+ "architectures": [
45
+ "GR00T_N1_5_RKD"
46
+ ],
47
+ "attn_implementation": null,
48
+ "backbone_cfg": {
49
+ "eagle_path": "NVEagle/eagle_er-qwen3_1_7B-Siglip2_400M_stage1_5_128gpu_er_v7_1mlp_nops",
50
+ "load_bf16": false,
51
+ "project_to_dim": null,
52
+ "reproject_vision": false,
53
+ "select_layer": 12,
54
+ "tune_llm": false,
55
+ "tune_visual": true,
56
+ "use_flash_attention": true
57
+ },
58
+ "compute_dtype": "bfloat16",
59
+ "hidden_size": 2048,
60
+ "model_dtype": "float32",
61
+ "model_type": "gr00t_n1_5",
62
+ "rkd_action_encoder_cls_ffn_hidden_dim": 2048,
63
+ "rkd_action_encoder_cls_num_heads": 8,
64
+ "rkd_action_encoder_cls_num_layers": 1,
65
+ "rkd_action_encoder_projector_dim": 512,
66
+ "rkd_action_encoder_projector_enabled": false,
67
+ "rkd_action_encoder_projector_pooling": "flatten",
68
+ "rkd_action_temp": 0.1,
69
+ "rkd_angle_loss_weight": 2.0,
70
+ "rkd_distance_loss_weight": 1.0,
71
+ "rkd_dtw_adaptive_floor": 0.05,
72
+ "rkd_dtw_adaptive_mix_lambda": 0.5,
73
+ "rkd_dtw_base_weight": 1.0,
74
+ "rkd_dtw_gripper_weight": 0.5,
75
+ "rkd_dtw_mix_mode": "distance_mix",
76
+ "rkd_dtw_normalize_components": true,
77
+ "rkd_dtw_pos_weight": 1.0,
78
+ "rkd_dtw_rot_weight": 1.0,
79
+ "rkd_dtw_timing_log_interval": 1000,
80
+ "rkd_enabled": true,
81
+ "rkd_exclude_diagonal": true,
82
+ "rkd_fm_loss_weight": 1.0,
83
+ "rkd_loss_type": "distance_angle",
84
+ "rkd_loss_weight": 0.5,
85
+ "rkd_loss_weight_end": 0.0,
86
+ "rkd_loss_weight_schedule": null,
87
+ "rkd_loss_weight_start": 0.1,
88
+ "rkd_raw_action_projector_dim": 512,
89
+ "rkd_raw_action_projector_enabled": false,
90
+ "rkd_raw_action_projector_pooling": "flatten",
91
+ "rkd_relation_mode": "flatten",
92
+ "rkd_student_reconstruction_head_dim": 512,
93
+ "rkd_student_reconstruction_head_enabled": true,
94
+ "rkd_student_reconstruction_head_hidden_dim": 512,
95
+ "rkd_student_reconstruction_head_pooling": "attention",
96
+ "rkd_teacher_source": "timewarp_action",
97
+ "rkd_timewarp_vae_checkpoint": "${HOME}/groot_robocasa/TimeWarping/TimewarpVAE/atomic65_z8_g16_h32_tw32_reg0p01_gb_ctrl_30k/checkpoint-step-30000-epoch-2.73",
98
+ "rkd_vlm_temp": 0.1,
99
+ "torch_dtype": "bfloat16",
100
+ "transformers_version": "4.51.3"
101
+ }
RKDv3/base120_p2-60k_p3-60k_fm1-rkd0p5_atomic65_human/default/phase3/experiment_cfg/metadata.json ADDED
@@ -0,0 +1,431 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "new_embodiment": {
3
+ "statistics": {
4
+ "state": {
5
+ "base_position": {
6
+ "max": [
7
+ 7.778976917266846,
8
+ 0.4876587688922882,
9
+ 0.7223060727119446
10
+ ],
11
+ "min": [
12
+ -4.94293737411499,
13
+ -6.890198230743408,
14
+ 0.6994727253913879
15
+ ],
16
+ "mean": [
17
+ 2.645730005800653,
18
+ -2.110066447108802,
19
+ 0.7004254938507055
20
+ ],
21
+ "std": [
22
+ 1.674798582069234,
23
+ 1.350392482318102,
24
+ 0.001107956680671985
25
+ ],
26
+ "q01": [
27
+ -1.0132281893317567,
28
+ -5.366714195503661,
29
+ 0.699973412653353
30
+ ],
31
+ "q99": [
32
+ 6.099414437993715,
33
+ -0.5498886012762193,
34
+ 0.7032860957123939
35
+ ]
36
+ },
37
+ "base_rotation": {
38
+ "max": [
39
+ 0.0,
40
+ 0.0,
41
+ 1.0,
42
+ 1.0
43
+ ],
44
+ "min": [
45
+ 0.0,
46
+ 0.0,
47
+ -1.0,
48
+ 0.0
49
+ ],
50
+ "mean": [
51
+ 0.0,
52
+ 0.0,
53
+ 0.24103797327895668,
54
+ 0.6071393324239354
55
+ ],
56
+ "std": [
57
+ 0.0,
58
+ 0.0,
59
+ 0.6615962784298529,
60
+ 0.36814045893518177
61
+ ],
62
+ "q01": [
63
+ 0.0,
64
+ 0.0,
65
+ -1.0,
66
+ 3.3913522062507517e-06
67
+ ],
68
+ "q99": [
69
+ 0.0,
70
+ 0.0,
71
+ 1.0,
72
+ 1.0
73
+ ]
74
+ },
75
+ "end_effector_position_relative": {
76
+ "max": [
77
+ 0.9014528393745422,
78
+ 0.8569459915161133,
79
+ 1.053884744644165
80
+ ],
81
+ "min": [
82
+ -0.5828549861907959,
83
+ -0.8486236333847046,
84
+ -0.3446122705936432
85
+ ],
86
+ "mean": [
87
+ 0.2641355221895152,
88
+ -0.01768191244102862,
89
+ 0.4458847037294584
90
+ ],
91
+ "std": [
92
+ 0.16437879912713016,
93
+ 0.24835923712456218,
94
+ 0.2124190803465676
95
+ ],
96
+ "q01": [
97
+ -0.004962867024380738,
98
+ -0.46862868595225443,
99
+ 0.20316236863116172
100
+ ],
101
+ "q99": [
102
+ 0.5495510296644974,
103
+ 0.44201455499146414,
104
+ 0.6984316542098391
105
+ ]
106
+ },
107
+ "end_effector_rotation_relative": {
108
+ "max": [
109
+ 0.9999998807907104,
110
+ 0.9997354745864868,
111
+ 0.9900747537612915,
112
+ 0.9735639095306396
113
+ ],
114
+ "min": [
115
+ -0.9999930262565613,
116
+ -0.9999884366989136,
117
+ -0.9975369572639465,
118
+ 6.754255821306288e-08
119
+ ],
120
+ "mean": [
121
+ -0.23536101367097556,
122
+ 0.0384247508727208,
123
+ -0.06854891687606551,
124
+ 0.1400254132624415
125
+ ],
126
+ "std": [
127
+ 0.8296216819146821,
128
+ 0.30067208350163366,
129
+ 0.33390768096919515,
130
+ 0.16926885514344236
131
+ ],
132
+ "q01": [
133
+ -0.9950540525525874,
134
+ -0.5590028004550295,
135
+ -0.5635208777456633,
136
+ 0.001747048046951923
137
+ ],
138
+ "q99": [
139
+ 0.9472459325196646,
140
+ 0.6183763159959191,
141
+ 0.5108060516185887,
142
+ 0.46646525101830405
143
+ ]
144
+ },
145
+ "gripper_qpos": {
146
+ "max": [
147
+ 0.055869169533252716,
148
+ 0.010916369967162609
149
+ ],
150
+ "min": [
151
+ -0.011669117026031017,
152
+ -0.06307756155729294
153
+ ],
154
+ "mean": [
155
+ 0.031849369408926076,
156
+ -0.03200748595341724
157
+ ],
158
+ "std": [
159
+ 0.012248040928874927,
160
+ 0.01206276030599065
161
+ ],
162
+ "q01": [
163
+ 0.006329614505051435,
164
+ -0.0405872199089872
165
+ ],
166
+ "q99": [
167
+ 0.040445365520021676,
168
+ -0.006282430560590391
169
+ ]
170
+ }
171
+ },
172
+ "action": {
173
+ "base_motion": {
174
+ "max": [
175
+ 1.0,
176
+ 1.0,
177
+ 1.0,
178
+ 0.0
179
+ ],
180
+ "min": [
181
+ -1.0,
182
+ -1.0,
183
+ -1.0,
184
+ 0.0
185
+ ],
186
+ "mean": [
187
+ 0.0071452133647496785,
188
+ 0.0015603604277540566,
189
+ -0.0014052511062474704,
190
+ 0.0
191
+ ],
192
+ "std": [
193
+ 0.11471999232420287,
194
+ 0.10209228015027566,
195
+ 0.0855718789814838,
196
+ 0.0
197
+ ],
198
+ "q01": [
199
+ -0.13329522793794643,
200
+ -0.16270717101832055,
201
+ -0.12619798329427676,
202
+ 0.0
203
+ ],
204
+ "q99": [
205
+ 0.1721366453818688,
206
+ 0.1972873452547982,
207
+ 0.11128571706089413,
208
+ 0.0
209
+ ]
210
+ },
211
+ "control_mode": {
212
+ "max": [
213
+ 1.0
214
+ ],
215
+ "min": [
216
+ -1.0
217
+ ],
218
+ "mean": [
219
+ -0.9218734936549906
220
+ ],
221
+ "std": [
222
+ 0.3874975703570591
223
+ ],
224
+ "q01": [
225
+ -1.0
226
+ ],
227
+ "q99": [
228
+ -0.26692269247002653
229
+ ]
230
+ },
231
+ "end_effector_position": {
232
+ "max": [
233
+ 1.0,
234
+ 1.0,
235
+ 1.0
236
+ ],
237
+ "min": [
238
+ -1.0,
239
+ -1.0,
240
+ -1.0
241
+ ],
242
+ "mean": [
243
+ 0.002920238070268553,
244
+ -0.0060913649045883315,
245
+ -0.07322447393298565
246
+ ],
247
+ "std": [
248
+ 0.4482114094812734,
249
+ 0.41233706159751987,
250
+ 0.4006084454909327
251
+ ],
252
+ "q01": [
253
+ -0.934104737554819,
254
+ -0.9255792016234818,
255
+ -0.8085822192859831
256
+ ],
257
+ "q99": [
258
+ 0.7874185230421324,
259
+ 0.9004049161564378,
260
+ 0.8018332495119154
261
+ ]
262
+ },
263
+ "end_effector_rotation": {
264
+ "max": [
265
+ 1.0,
266
+ 1.0,
267
+ 1.0
268
+ ],
269
+ "min": [
270
+ -1.0,
271
+ -1.0,
272
+ -1.0
273
+ ],
274
+ "mean": [
275
+ 0.0077772595328458915,
276
+ -0.022620543966259083,
277
+ -0.00029332176634943376
278
+ ],
279
+ "std": [
280
+ 0.11155576533045976,
281
+ 0.13202104364511658,
282
+ 0.1240098754411023
283
+ ],
284
+ "q01": [
285
+ -0.2726786325766535,
286
+ -0.4125887853047325,
287
+ -0.3445183934868137
288
+ ],
289
+ "q99": [
290
+ 0.31803481867756295,
291
+ 0.3015153519871841,
292
+ 0.34730986869917346
293
+ ]
294
+ },
295
+ "gripper_close": {
296
+ "max": [
297
+ 1.0
298
+ ],
299
+ "min": [
300
+ -1.0
301
+ ],
302
+ "mean": [
303
+ -0.33422761717726496
304
+ ],
305
+ "std": [
306
+ 0.9425001619711008
307
+ ],
308
+ "q01": [
309
+ -1.0
310
+ ],
311
+ "q99": [
312
+ 0.681821585371573
313
+ ]
314
+ }
315
+ }
316
+ },
317
+ "modalities": {
318
+ "video": {
319
+ "robot0_eye_in_hand": {
320
+ "resolution": [
321
+ 256,
322
+ 256
323
+ ],
324
+ "channels": 3,
325
+ "fps": 20.0
326
+ },
327
+ "robot0_agentview_left": {
328
+ "resolution": [
329
+ 256,
330
+ 256
331
+ ],
332
+ "channels": 3,
333
+ "fps": 20.0
334
+ },
335
+ "robot0_agentview_right": {
336
+ "resolution": [
337
+ 256,
338
+ 256
339
+ ],
340
+ "channels": 3,
341
+ "fps": 20.0
342
+ }
343
+ },
344
+ "state": {
345
+ "base_position": {
346
+ "absolute": true,
347
+ "rotation_type": null,
348
+ "shape": [
349
+ 3
350
+ ],
351
+ "continuous": true
352
+ },
353
+ "base_rotation": {
354
+ "absolute": true,
355
+ "rotation_type": "quaternion",
356
+ "shape": [
357
+ 4
358
+ ],
359
+ "continuous": true
360
+ },
361
+ "end_effector_position_relative": {
362
+ "absolute": true,
363
+ "rotation_type": null,
364
+ "shape": [
365
+ 3
366
+ ],
367
+ "continuous": true
368
+ },
369
+ "end_effector_rotation_relative": {
370
+ "absolute": true,
371
+ "rotation_type": "quaternion",
372
+ "shape": [
373
+ 4
374
+ ],
375
+ "continuous": true
376
+ },
377
+ "gripper_qpos": {
378
+ "absolute": true,
379
+ "rotation_type": null,
380
+ "shape": [
381
+ 2
382
+ ],
383
+ "continuous": true
384
+ }
385
+ },
386
+ "action": {
387
+ "base_motion": {
388
+ "absolute": true,
389
+ "rotation_type": null,
390
+ "shape": [
391
+ 4
392
+ ],
393
+ "continuous": true
394
+ },
395
+ "control_mode": {
396
+ "absolute": true,
397
+ "rotation_type": null,
398
+ "shape": [
399
+ 1
400
+ ],
401
+ "continuous": true
402
+ },
403
+ "end_effector_position": {
404
+ "absolute": true,
405
+ "rotation_type": null,
406
+ "shape": [
407
+ 3
408
+ ],
409
+ "continuous": true
410
+ },
411
+ "end_effector_rotation": {
412
+ "absolute": true,
413
+ "rotation_type": "axis_angle",
414
+ "shape": [
415
+ 3
416
+ ],
417
+ "continuous": true
418
+ },
419
+ "gripper_close": {
420
+ "absolute": true,
421
+ "rotation_type": null,
422
+ "shape": [
423
+ 1
424
+ ],
425
+ "continuous": true
426
+ }
427
+ }
428
+ },
429
+ "embodiment_tag": "new_embodiment"
430
+ }
431
+ }
RKDv3/base120_p2-60k_p3-60k_fm1-rkd0p5_atomic65_human/default/phase3/model-00001-of-00002.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:970f2ebce9fab15bff554fe8c34e03bf7a41c6fded29e9cc87b9e88d949cd513
3
+ size 4999367032
RKDv3/base120_p2-60k_p3-60k_fm1-rkd0p5_atomic65_human/default/phase3/model-00002-of-00002.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:381bedadefe51967a50b34a0955d2b4419e636a6b5a01fa0efecbf7af537751a
3
+ size 2664461700
RKDv3/base120_p2-60k_p3-60k_fm1-rkd0p5_atomic65_human/default/phase3/model.safetensors.index.json ADDED
The diff for this file is too large to render. See raw diff
 
RKDv3/base120_p2-60k_p3-60k_fm1-rkd0p5_atomic65_human/default/phase3/trainer_state.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fb6283653dc62535bad21f34e7a02968a4dfa4ffc12aebf5fe785576651f6633
3
+ size 11204654
RKDv3/base120_p2-60k_p3-60k_fm1-rkd0p5_atomic65_human/default/phase3/training_args.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7f6adae4d167bebb21c718580489904d78057648f9406eb279c697d2620741d3
3
+ size 6033
RKDv3/base120_p2-60k_p3-60k_fm1-rkd0p5_atomic65_human/default/resolved_config.yaml ADDED
@@ -0,0 +1,50 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ experiment_name: RKDv3 TimewarpVAE DA, Student:APHead from GR00T base to 120k on Atomic65
2
+ human, phase2 60k + phase3 60k, FM:RKD 1:0.5
3
+ policy_type: groot_RKDv3
4
+ base_model_path: nvidia/GR00T-N1.5-3B
5
+ output_root: ${HOME}/groot_robocasa/robocasa_v2/RKDv3/base120_p2-60k_p3-60k_fm1-rkd0p5_atomic65_human
6
+ dataset:
7
+ dataset_soup: atomic_pretrain_human
8
+ training:
9
+ num_gpus: 1
10
+ batch_size: 128
11
+ seed: 42
12
+ model:
13
+ rkd_teacher_source: timewarp_action
14
+ rkd_timewarp_vae_checkpoint: ${HOME}/groot_robocasa/TimeWarping/TimewarpVAE/atomic65_z8_g16_h32_tw32_reg0p01_gb_ctrl_30k/checkpoint-step-30000-epoch-2.73
15
+ rkd_relation_mode: flatten
16
+ rkd_loss_type: distance_angle
17
+ rkd_student_reconstruction_head_pooling: attention
18
+ rkd_student_reconstruction_head_enabled: true
19
+ rkd_student_reconstruction_head_dim: 512
20
+ rkd_student_reconstruction_head_hidden_dim: 512
21
+ rkd_distance_loss_weight: 1.0
22
+ rkd_angle_loss_weight: 2.0
23
+ rkd_exclude_diagonal: true
24
+ phases:
25
+ - name: phase2_rkdv3_timewarp_vae_da_aphead_only_60k
26
+ max_steps: 60000
27
+ save_steps: 0
28
+ trainable:
29
+ preset: processing_line_only
30
+ tune_llm: false
31
+ tune_visual: false
32
+ tune_projector: false
33
+ tune_diffusion_model: false
34
+ losses:
35
+ rkd_enabled: true
36
+ rkd_fm_loss_weight: 0.0
37
+ rkd_loss_weight: 1.0
38
+ - name: phase3_fm_rkdv3_timewarp_vae_da_aphead_fixed_0p5_60k
39
+ max_steps: 60000
40
+ save_steps: 0
41
+ trainable:
42
+ tune_llm: false
43
+ tune_visual: false
44
+ tune_projector: true
45
+ tune_diffusion_model: true
46
+ losses:
47
+ rkd_enabled: true
48
+ rkd_fm_loss_weight: 1.0
49
+ rkd_loss_weight: 0.5
50
+ resolved_sweep: {}
RKDv3/base120_single_fm1-rkd1to0_cosine_atomic65_human/default/base120_single_fm1-rkd1to0_cosine_atomic65_human.yaml CHANGED
@@ -6,12 +6,12 @@ output_root: ${HOME}/groot_robocasa/robocasa_v2/RKDv3/base120_single_fm1-rkd1to0
6
  dataset:
7
  dataset_soup: atomic_pretrain_human
8
  training:
9
- num_gpus: 2
10
- batch_size: 64
11
  seed: 42
12
  model:
13
  rkd_teacher_source: timewarp_action
14
- rkd_timewarp_vae_checkpoint: ${HOME}/groot_robocasa/TimeWarping/TimewarpVAE/atomic65_z8_g16_h32_tw32_reg0p01_gb_ctrl_30k/checkpoint-step-30000-epoch-2.73
15
  rkd_relation_mode: flatten
16
  rkd_loss_type: distance_angle
17
  rkd_student_reconstruction_head_pooling: attention
 
6
  dataset:
7
  dataset_soup: atomic_pretrain_human
8
  training:
9
+ num_gpus: 1
10
+ batch_size: 128
11
  seed: 42
12
  model:
13
  rkd_teacher_source: timewarp_action
14
+ rkd_timewarp_vae_checkpoint: ${HOME}/groot_robocasa/TimeWarping/TimewarpVAE/robocasa_vanilla_t16_z8_g8_h32_reg0p01_beta0p01/checkpoint-step-60000-epoch-17.46
15
  rkd_relation_mode: flatten
16
  rkd_loss_type: distance_angle
17
  rkd_student_reconstruction_head_pooling: attention
RKDv3/base120_single_fm1-rkd1to0_cosine_atomic65_human/default/resolved_config.yaml CHANGED
@@ -6,12 +6,12 @@ output_root: ${HOME}/groot_robocasa/robocasa_v2/RKDv3/base120_single_fm1-rkd1to0
6
  dataset:
7
  dataset_soup: atomic_pretrain_human
8
  training:
9
- num_gpus: 2
10
- batch_size: 64
11
  seed: 42
12
  model:
13
  rkd_teacher_source: timewarp_action
14
- rkd_timewarp_vae_checkpoint: ${HOME}/groot_robocasa/TimeWarping/TimewarpVAE/atomic65_z8_g16_h32_tw32_reg0p01_gb_ctrl_30k/checkpoint-step-30000-epoch-2.73
15
  rkd_relation_mode: flatten
16
  rkd_loss_type: distance_angle
17
  rkd_student_reconstruction_head_pooling: attention
 
6
  dataset:
7
  dataset_soup: atomic_pretrain_human
8
  training:
9
+ num_gpus: 1
10
+ batch_size: 128
11
  seed: 42
12
  model:
13
  rkd_teacher_source: timewarp_action
14
+ rkd_timewarp_vae_checkpoint: ${HOME}/groot_robocasa/TimeWarping/TimewarpVAE/robocasa_vanilla_t16_z8_g8_h32_reg0p01_beta0p01/checkpoint-step-60000-epoch-17.46
15
  rkd_relation_mode: flatten
16
  rkd_loss_type: distance_angle
17
  rkd_student_reconstruction_head_pooling: attention