| { |
| "suite": "level2_from_level3", |
| "method": "PI0.5 LoRA + trainable weights, multi-task; level-to-level transfer", |
| "source_level": "level_3", |
| "target_level": "level_2", |
| "init_checkpoint": "multi_task/level_3/lora_step_0080000.npz", |
| "init_note": "All weights (base + LoRA + trainable towers) initialised from the level_3 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 (50 -> 44): 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": "08c362b52e8c46becce077a8924ac75a664a77fdf15035d3f359fa5800491945", |
| "checkpoint_bytes": 1867943218, |
| "tensor_count": 51, |
| "trainable_params": 466981660, |
| "tasks": [ |
| "HangMug-v1", |
| "PivotBook-v1", |
| "OpenBoxRetrieval-v1", |
| "KitAssembly-v1", |
| "GraspTwoBlocks-v1", |
| "ReorientPen-v1", |
| "DexOpenDoor-v1", |
| "HammerNailOut-v1", |
| "DualArmLiftBall-v1", |
| "DualArmBookPickPlace-v1", |
| "DualArmPegInHole-v1", |
| "DualArmCleanupPotEnv-v1", |
| "DualArmOpenLaptop-v1", |
| "DualArmThrowBlock-v1", |
| "DualArmJugglingBall-v1", |
| "DualArmOpenBottle-v1", |
| "DualArmPlaceRopeOverRod-v1", |
| "DualArmStuffCloth-v1", |
| "DualArmPouringTwoCups-v1", |
| "DualArmClothAlign-v1", |
| "DualArmPushObject-v1", |
| "DualArmStraightenRope-v1", |
| "ClothWipe-v1", |
| "RollingPin-v1", |
| "RopeBridgeDex-v1", |
| "FoldShirt-v1", |
| "ScoopWater-v1", |
| "StraightenTwoRopesSequential-v1" |
| ], |
| "num_tasks": 28, |
| "prompts": { |
| "HangMug-v1": "perform the ManiLadder task: hang mug", |
| "PivotBook-v1": "perform the ManiLadder task: pivot book", |
| "OpenBoxRetrieval-v1": "perform the ManiLadder task: open box retrieval", |
| "KitAssembly-v1": "perform the ManiLadder task: kit assembly", |
| "GraspTwoBlocks-v1": "perform the ManiLadder task: grasp two blocks", |
| "ReorientPen-v1": "perform the ManiLadder task: reorient pen", |
| "DexOpenDoor-v1": "perform the ManiLadder task: dex open door", |
| "HammerNailOut-v1": "perform the ManiLadder task: hammer nail out", |
| "DualArmLiftBall-v1": "perform the ManiLadder task: dual arm lift ball", |
| "DualArmBookPickPlace-v1": "perform the ManiLadder task: dual arm book pick place", |
| "DualArmPegInHole-v1": "perform the ManiLadder task: dual arm peg in hole", |
| "DualArmCleanupPotEnv-v1": "perform the ManiLadder task: dual arm cleanup pot env", |
| "DualArmOpenLaptop-v1": "perform the ManiLadder task: dual arm open laptop", |
| "DualArmThrowBlock-v1": "perform the ManiLadder task: dual arm throw block", |
| "DualArmJugglingBall-v1": "perform the ManiLadder task: dual arm juggling ball", |
| "DualArmOpenBottle-v1": "perform the ManiLadder task: dual arm open bottle", |
| "DualArmPlaceRopeOverRod-v1": "perform the ManiLadder task: dual arm place rope over rod", |
| "DualArmStuffCloth-v1": "perform the ManiLadder task: dual arm stuff cloth", |
| "DualArmPouringTwoCups-v1": "perform the ManiLadder task: dual arm pouring two cups", |
| "DualArmClothAlign-v1": "perform the ManiLadder task: dual arm cloth align", |
| "DualArmPushObject-v1": "perform the ManiLadder task: dual arm push object", |
| "DualArmStraightenRope-v1": "perform the ManiLadder task: dual arm straighten rope", |
| "ClothWipe-v1": "perform the ManiLadder task: cloth wipe", |
| "RollingPin-v1": "perform the ManiLadder task: rolling pin", |
| "RopeBridgeDex-v1": "perform the ManiLadder task: rope bridge dex", |
| "FoldShirt-v1": "perform the ManiLadder task: fold shirt", |
| "ScoopWater-v1": "perform the ManiLadder task: scoop water", |
| "StraightenTwoRopesSequential-v1": "perform the ManiLadder task: straighten two ropes sequential" |
| }, |
| "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": "dd578f246b55ab92eff2f93926ef0d006826ba53fbe4fa1d4e592c0f4d28af45", |
| "norm_stats_note": "identical to multi_task/level_2/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_2: Linslab/ManiLadder_Dataset raw/<task>, 20 demos per task, revision 1190737391cc", |
| "training_data_overrides": {}, |
| "excluded_tasks": {}, |
| "eval_episodes": 50, |
| "eval_seed_base": 1000, |
| "evaluated_steps": [ |
| 50000 |
| ], |
| "campaign": "pi05_transfer_20261101", |
| "trained": "2026-11-01 04:33 -> 18:27 (lyg2026)" |
| } |
|
|
