Linslab's picture
Add level_4 PI0.5 multi-task LoRA (27 tasks, step 80000, 26/27 tasks evaluated, mean 0.001 over 50 ep)
491b15c verified
Raw History Blame Contribute Delete
4.8 kB
{
"suite": "level_4",
"method": "PI0.5 LoRA + trainable weights, multi-task (one policy for every task of the level)",
"base_checkpoint": "gs://openpi-assets/checkpoints/pi05_base/params",
"checkpoint": "lora_step_0080000.npz",
"checkpoint_sha256": "6282d8278bfbc4c9bcefd2f520108fa162113fd6a6713d6cf398b86ce70541c4",
"checkpoint_bytes": 1867992394,
"tensor_count": 51,
"trainable_params": 466993954,
"tasks": [
"RingTossing-v1",
"HitBilliard-v1",
"FlipPan-v1",
"PingPongBallTransport-v1",
"FitnessBall-v1",
"TennisRally-v1",
"RodThroughHoles-v1",
"SpinPenHora-v1",
"DualArmBallClassify-v1",
"DualArmBilliardBridge-v1",
"DualArmRubikCube-v1",
"DualArmMahjongSetup-v1",
"DualArmKeyboardType-v1",
"DualArmSendPingpong-v1",
"DualArmFlipTable-v1",
"DualArmShootBasketball-v1",
"DualArmClothCylinderWipe-v1",
"DualArmJumpRope-v1",
"DualArmLiquidMixing-v1",
"DualArmClothHanger-v1",
"DualArmSandShovel-v1",
"DualArmWallRings-v1",
"CoverPattern-v1",
"ClaySealStamp-v1",
"BuildTent-v1",
"WrapRope-v1",
"StampRoll-v1"
],
"num_tasks": 27,
"prompts": {
"RingTossing-v1": "perform the ManiLadder task: ring tossing",
"HitBilliard-v1": "perform the ManiLadder task: hit billiard",
"FlipPan-v1": "perform the ManiLadder task: flip pan",
"PingPongBallTransport-v1": "perform the ManiLadder task: ping pong ball transport",
"FitnessBall-v1": "perform the ManiLadder task: fitness ball",
"TennisRally-v1": "perform the ManiLadder task: tennis rally",
"RodThroughHoles-v1": "perform the ManiLadder task: rod through holes",
"SpinPenHora-v1": "perform the ManiLadder task: spin pen hora",
"DualArmBallClassify-v1": "perform the ManiLadder task: dual arm ball classify",
"DualArmBilliardBridge-v1": "perform the ManiLadder task: dual arm billiard bridge",
"DualArmRubikCube-v1": "perform the ManiLadder task: dual arm rubik cube",
"DualArmMahjongSetup-v1": "perform the ManiLadder task: dual arm mahjong setup",
"DualArmKeyboardType-v1": "perform the ManiLadder task: dual arm keyboard type",
"DualArmSendPingpong-v1": "perform the ManiLadder task: dual arm send pingpong",
"DualArmFlipTable-v1": "perform the ManiLadder task: dual arm flip table",
"DualArmShootBasketball-v1": "perform the ManiLadder task: dual arm shoot basketball",
"DualArmClothCylinderWipe-v1": "perform the ManiLadder task: dual arm cloth cylinder wipe",
"DualArmJumpRope-v1": "perform the ManiLadder task: dual arm jump rope",
"DualArmLiquidMixing-v1": "perform the ManiLadder task: dual arm liquid mixing",
"DualArmClothHanger-v1": "perform the ManiLadder task: dual arm cloth hanger",
"DualArmSandShovel-v1": "perform the ManiLadder task: dual arm sand shovel",
"DualArmWallRings-v1": "perform the ManiLadder task: dual arm wall rings",
"CoverPattern-v1": "perform the ManiLadder task: cover pattern",
"ClaySealStamp-v1": "perform the ManiLadder task: clay seal stamp",
"BuildTent-v1": "perform the ManiLadder task: build tent",
"WrapRope-v1": "perform the ManiLadder task: wrap rope",
"StampRoll-v1": "perform the ManiLadder task: stamp roll"
},
"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": 80000,
"batch_size": 64,
"lr": 5e-05,
"warmup_steps": 1000,
"lr_schedule": "cosine decay over all 80k steps",
"data_parallel_gpus": 8,
"gpu": "NVIDIA A100 80GB",
"fsdp_devices": 1,
"online_eval": false,
"norm_stats_file": "norm_stats.json",
"norm_stats_sha256": "7a96ea78deee9779f6ca47b9e92d178dffa94f346760bdc2644f951828d51d65",
"norm_stats_note": "quantile stats computed per dimension while ignoring padded (all-zero) action/state dimensions of embodiments narrower than action_dim, so q01==q99 collapse does not blow up scaling",
"training_data": "Linslab/ManiLadder_Dataset raw/<task>, 20 demos per task, revision 1190737391cc (sha256-verified download 2026-10-28)",
"training_data_overrides": {},
"excluded_tasks": {
"SwingRing-v1": "level 4, excluded on request",
"OpenCoffer-v1": "no demos exist",
"DualArmSqueezeClayDex-v1": "no demos exist",
"DualArmRopeScissorCut-v1": "no demos exist",
"DualArmBuildCanel-v1": "no demos exist",
"PickSoft-v1": "no demos exist",
"ClayMolding-v1": "no demos exist"
},
"eval_episodes": 50,
"eval_seed_base": 1000,
"evaluated_steps": [
80000
],
"campaign": "pi05_levels_20261028",
"trained": "2026-10-29 02:10 -> 2026-10-30 (full restart with padding-aware norm stats)"
}