actaug_prio_1024
Action-augmented RoboCasa PickPlaceCounterToCabinet demonstrations converted to the
GR00T-N1.5 / LeRobot v2.1 layout (gr00t_views). 1024 episodes, 128 per target object.
Source: robocasa_aug1024_prio_curated. Episode order is
sorted(os.listdir(source)) over directories starting with ep; index/episode_index.csv
records the exact mapping from episode_index to source episode, verified by matching
every episode length (1024/1024).
Contents
| episodes | 1024 |
| frames | 248,524 |
| per object | 128 × 8 |
| video keys | 3 (ego_view, exterior_image_1, wrist_image) |
| state dim | 16 |
| action dim | 12 |
| size | 640 MB |
| 1 epoch @ global batch 64 | 3,883 steps |
Target objects: steak/steak_8, teapot/teapot_6, teapot/teapot_7,
donut/donut_5, jar/Jar023, measuring_cup/MeasuringCup009,
soap_dispenser/SoapDispenser010, syrup_bottle/SyrupBottle005.
Index files
index/ carries the provenance for this dataset and for the two disjoint
512-episode subsamples drawn from the same curated pool:
| file | rows |
|---|---|
index/actaug_prio_1024/episode_index.{csv,jsonl} |
1024 |
index/actaug_prio_512/episode_index.{csv,jsonl} |
512 — first 64 per object (sample A) |
index/actaug_prio_512b/episode_index.{csv,jsonl} |
512 — last 64 per object (sample B) |
Columns: episode_index, source_episode, source_demo, category, model, seed, length, task.
Samples A and B share no episodes.
Action layout
Actions follow the LeRobot order declared in meta/modality.json:
base_motion[0:4] control_mode[4:5] eef_pos[5:8] eef_rot[8:11] gripper_close[11:12]
This differs from the simulator's own order (eef_pos, eef_rot, gripper, base, torso, base_mode). Converting a RoboCasa HDF5 without the permutation
[7,8,9,10,11,0,1,2,3,4,5,6] yields a dataset that trains to 0% success. See
dataset/repair_action_order.py in
mlnha/pose6daug-scripts.
Eval results
Fine-tuned from a 60k-step RoboCasa base checkpoint, global batch 64, evaluated by exact-state replay over 160 held-out episodes (20 per object), policy seed 12345.
| checkpoint | epochs | success | grasp | lift | in_cab |
|---|---|---|---|---|---|
| 1500 | 0.39 | 44/160 = 27.5% | 64 | 59 | 50 |
| 3000 | 0.77 | 25/160 = 15.6% | 68 | 36 | 32 |
| 5000 | 1.29 | 25/160 = 15.6% | 71 | 37 | 34 |
Baseline (no augmentation, same protocol): 11/160 = 6.9%.
Grasping does not degrade with longer training (64 → 68 → 71); the lift stage does (59 → 36 → 37). The peak sits at 0.39 epoch.
Augmentation, conversion, training and evaluation code: mlnha/pose6daug-scripts.
- Downloads last month
- 310