File size: 5,305 Bytes
c4568d7 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 | {
"suite": "level3_from_level2",
"method": "PI0.5 LoRA + trainable weights, multi-task; level-to-level transfer",
"source_level": "level_2",
"target_level": "level_3",
"init_checkpoint": "multi_task/level_2/lora_step_0080000.npz",
"init_note": "All weights (base + LoRA + trainable towers) initialised from the level_2 multi-task policy at step 80000 (its full orbax params); the optimizer, warmup and cosine schedule were restarted. action_dim differs between source and target (44 -> 50): action_in_proj/kernel, action_out_proj/kernel and action_out_proj/bias were copied on the overlapping slice and the new slots initialised like a fresh Flax Linear layer; every other tensor was copied as is.",
"base_checkpoint": "gs://openpi-assets/checkpoints/pi05_base/params (via the source level policy)",
"checkpoint": "lora_step_0050000.npz",
"checkpoint_sha256": "e70b239f67df087c387c75c92c3a6251cf8cea7b2f5ca684b3d55587f8604045",
"checkpoint_bytes": 1867992394,
"tensor_count": 51,
"trainable_params": 466993954,
"tasks": [
"GroceryPrepare-v1",
"Stack6Blocks-v1",
"PushTargetPose-v1",
"MazeBlockPush-v1",
"TweezersPickPlace-v1",
"ClutterGrasp-v1",
"MobilePickPlace-v1",
"HammerNail-v1",
"DualArmFoldCarton-v1",
"DualArmOpenFridge-v1",
"DualArmAssembleBracket-v1",
"DualArmOrganizeCabinet-v1",
"DualArmPingpongBalance-v1",
"DualArmSetupTable-v1",
"DualArmMakeLine-v1",
"DualArmPlugCharger-v1",
"DualArmLayerClayInDish-v1",
"DualArmThreadCloth-v1",
"DualArmThreading-v1",
"DualArmWeaveCloth-v1",
"DualArmShapeDoll-v1",
"DualArmThreadRopeThroughTube-v1",
"HitBox-v1",
"PressFitSoftPeg-v1",
"LiquidMeasuring-v1",
"SeparateTwoRopesEnv-v1"
],
"num_tasks": 26,
"prompts": {
"GroceryPrepare-v1": "perform the ManiLadder task: grocery prepare",
"Stack6Blocks-v1": "perform the ManiLadder task: stack6 blocks",
"PushTargetPose-v1": "perform the ManiLadder task: push target pose",
"MazeBlockPush-v1": "perform the ManiLadder task: maze block push",
"TweezersPickPlace-v1": "perform the ManiLadder task: tweezers pick place",
"ClutterGrasp-v1": "perform the ManiLadder task: clutter grasp",
"MobilePickPlace-v1": "perform the ManiLadder task: mobile pick place",
"HammerNail-v1": "perform the ManiLadder task: hammer nail",
"DualArmFoldCarton-v1": "perform the ManiLadder task: dual arm fold carton",
"DualArmOpenFridge-v1": "perform the ManiLadder task: dual arm open fridge",
"DualArmAssembleBracket-v1": "perform the ManiLadder task: dual arm assemble bracket",
"DualArmOrganizeCabinet-v1": "perform the ManiLadder task: dual arm organize cabinet",
"DualArmPingpongBalance-v1": "perform the ManiLadder task: dual arm pingpong balance",
"DualArmSetupTable-v1": "perform the ManiLadder task: dual arm setup table",
"DualArmMakeLine-v1": "perform the ManiLadder task: dual arm make line",
"DualArmPlugCharger-v1": "perform the ManiLadder task: dual arm plug charger",
"DualArmLayerClayInDish-v1": "perform the ManiLadder task: dual arm layer clay in dish",
"DualArmThreadCloth-v1": "perform the ManiLadder task: dual arm thread cloth",
"DualArmThreading-v1": "perform the ManiLadder task: dual arm threading",
"DualArmWeaveCloth-v1": "perform the ManiLadder task: dual arm weave cloth",
"DualArmShapeDoll-v1": "perform the ManiLadder task: dual arm shape doll",
"DualArmThreadRopeThroughTube-v1": "perform the ManiLadder task: dual arm thread rope through tube",
"HitBox-v1": "perform the ManiLadder task: hit box",
"PressFitSoftPeg-v1": "perform the ManiLadder task: press fit soft peg",
"LiquidMeasuring-v1": "perform the ManiLadder task: liquid measuring",
"SeparateTwoRopesEnv-v1": "perform the ManiLadder task: separate two ropes env"
},
"num_demos_per_task": 20,
"action_dim": 50,
"state_dim": 50,
"state_input": "zero vector",
"action_horizon": 50,
"max_token_len": 200,
"image_size": 224,
"exec_steps": 16,
"sample_steps": 10,
"control_mode": "pd_ee_delta_pose",
"train_steps": 50000,
"batch_size": 64,
"lr": 5e-05,
"warmup_steps": 1000,
"lr_schedule": "cosine decay 5e-5 -> 5e-6 over all 50000 steps",
"data_parallel_gpus": 8,
"gpu": "NVIDIA A100 80GB",
"fsdp_devices": 1,
"online_eval": false,
"norm_stats_file": "norm_stats.json",
"norm_stats_sha256": "a56198ac52c8710376ba23fb17460d6ffee6e7a2c4d90b2e2f98b3d319e3e4d4",
"norm_stats_note": "identical to multi_task/level_3/norm_stats.json: quantile stats of the target level, computed per dimension while ignoring padded (all-zero) action/state dimensions",
"training_data": "same dataset as multi_task/level_3: Linslab/ManiLadder_Dataset raw/<task>, 20 demos per task, revision 1190737391cc",
"training_data_overrides": {
"MazeBlockPush-v1": "local demos collected 2026-09-09 after the layout-pool change (HF copy of 2026-08-28 predates the env)"
},
"excluded_tasks": {
"BagHanging-v1": "level 3, excluded on request",
"BagDropIn-v1": "level 3, excluded on request"
},
"eval_episodes": 50,
"eval_seed_base": 1000,
"evaluated_steps": [
50000
],
"campaign": "pi05_transfer_20261101",
"trained": "2026-11-01 04:20 -> 18:17 (lyg2056)"
}
|