File size: 5,052 Bytes
42660db | 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 99 100 101 102 103 104 105 106 107 108 109 110 | {
"suite": "level1_from_level2",
"method": "PI0.5 LoRA + trainable weights, multi-task; level-to-level transfer",
"source_level": "level_2",
"target_level": "level_1",
"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 identical (44); every tensor copied as is.",
"base_checkpoint": "gs://openpi-assets/checkpoints/pi05_base/params (via the source level policy)",
"checkpoint": "lora_step_0050000.npz",
"checkpoint_sha256": "d908cac219a367be30aeb3546018a3555702a66b803efd659ce2ef67e36d8da8",
"checkpoint_bytes": 1867943218,
"tensor_count": 51,
"trainable_params": 466981660,
"tasks": [
"PushSoccer-v1",
"PickPlacePen-v1",
"OpenDoor-v1",
"StackBlock-v1",
"LiftBucket-v1",
"TurnOnFaucet-v1",
"ReorientCube-v1",
"GraspCylinder-v1",
"DualArmPushBox-v1",
"DualArmDoubleDoorBox-v1",
"DualArmButtonPress-v1",
"DualArmHammerHandover-v1",
"DualArmPressButton-v1",
"DualArmOpenScissor-v1",
"DualArmSwingCup-v1",
"DualArmOpenPen-v1",
"DualArmCoverCloth-v1",
"DualArmLiftSoftDex-v1",
"DualArmRopeStretch-v1",
"DualArmFoldHandKerchief-v1",
"DualArmPouringWater-v1",
"DualArmRopeCylinderWarpping-v1",
"DrapeCloth-v1",
"RopePalmCageHoldDex-v1",
"SqueezeClay-v1",
"ShapeTToI-v1",
"StraightenRope-v1",
"UnfoldShirt-v1"
],
"num_tasks": 28,
"prompts": {
"PushSoccer-v1": "perform the ManiLadder task: push soccer",
"PickPlacePen-v1": "perform the ManiLadder task: pick place pen",
"OpenDoor-v1": "perform the ManiLadder task: open door",
"StackBlock-v1": "perform the ManiLadder task: stack block",
"LiftBucket-v1": "perform the ManiLadder task: lift bucket",
"TurnOnFaucet-v1": "perform the ManiLadder task: turn on faucet",
"ReorientCube-v1": "perform the ManiLadder task: reorient cube",
"GraspCylinder-v1": "perform the ManiLadder task: grasp cylinder",
"DualArmPushBox-v1": "perform the ManiLadder task: dual arm push box",
"DualArmDoubleDoorBox-v1": "perform the ManiLadder task: dual arm double door box",
"DualArmButtonPress-v1": "perform the ManiLadder task: dual arm button press",
"DualArmHammerHandover-v1": "perform the ManiLadder task: dual arm hammer handover",
"DualArmPressButton-v1": "perform the ManiLadder task: dual arm press button",
"DualArmOpenScissor-v1": "perform the ManiLadder task: dual arm open scissor",
"DualArmSwingCup-v1": "perform the ManiLadder task: dual arm swing cup",
"DualArmOpenPen-v1": "perform the ManiLadder task: dual arm open pen",
"DualArmCoverCloth-v1": "perform the ManiLadder task: dual arm cover cloth",
"DualArmLiftSoftDex-v1": "perform the ManiLadder task: dual arm lift soft dex",
"DualArmRopeStretch-v1": "perform the ManiLadder task: dual arm rope stretch",
"DualArmFoldHandKerchief-v1": "perform the ManiLadder task: dual arm fold hand kerchief",
"DualArmPouringWater-v1": "perform the ManiLadder task: dual arm pouring water",
"DualArmRopeCylinderWarpping-v1": "perform the ManiLadder task: dual arm rope cylinder warpping",
"DrapeCloth-v1": "perform the ManiLadder task: drape cloth",
"RopePalmCageHoldDex-v1": "perform the ManiLadder task: rope palm cage hold dex",
"SqueezeClay-v1": "perform the ManiLadder task: squeeze clay",
"ShapeTToI-v1": "perform the ManiLadder task: shape t to i",
"StraightenRope-v1": "perform the ManiLadder task: straighten rope",
"UnfoldShirt-v1": "perform the ManiLadder task: unfold shirt"
},
"num_demos_per_task": 20,
"action_dim": 44,
"state_dim": 44,
"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": "2ee8037535393988cf87ad99254ae9b551e025a441d7eab261139ec9ca6393f2",
"norm_stats_note": "identical to multi_task/level_1/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_1: Linslab/ManiLadder_Dataset raw/<task>, 20 demos per task, revision 1190737391cc",
"training_data_overrides": {
"DualArmOpenScissor-v1": "local demos collected 2026-10-22 (HF copy of 2026-06-11 predates the env)"
},
"excluded_tasks": {},
"eval_episodes": 50,
"eval_seed_base": 1000,
"evaluated_steps": [
50000
],
"campaign": "pi05_transfer_20261101",
"trained": "2026-11-01 04:20 -> 18:28 (lyg2122)"
}
|