Piston Job A N1.6 Bimanual 3-Camera 40K
This repository contains the model-only 40,000-step checkpoint and the exact
training/evaluation recipe for Piston Job A, the bimanual task. It is not
the Job B piston-drop/press checkpoint.
Task
Verbatim training instruction:
pick up the piston with the right hand, grasp the tube with the left hand,
release and regrasp the piston, then bring the piston and tube together.
Job A starts with the Tube on the desktop and both hands open. The task uses
both arms and both Inspire hands, includes an intentional right-hand release
and regrasp, and ends with Piston/Tube docking. It does not use Job B's
hand-held Tube reset or legacy two-press objective.
Training data contract
- Source:
g1_inspire_piston_bi_manual_3cam_lerobot_v21_20260819
- Semantic-success episodes: 87
- Frames: 90,275 at 50 FPS
- Cameras:
ego_view, left_wrist_view, right_wrist_view
- State: 29 dimensions (
left_arm, right_arm, left_hand, right_hand, waist)
- Action: 30 dimensions
- Arms: relative joint actions
- Hands: absolute raw Isaac/USD Inspire actions
- Action delta offsets:
0..29 (30 decoded serving steps)
The immutable source hashes and validated input contract are included under
recipe/contracts/.
Training recipe
- Base: NVIDIA GR00T N1.6 3B, revision
d0814e7ecb19202e7c8468b46098b0b7ef3a6d61
- Steps: 40,000
- Micro batch: 4
- Gradient accumulation: 8
- Effective batch: 32
- Learning rate:
1e-4, cosine schedule, 5% warmup
- Weight decay:
1e-5
- State dropout:
0.8
- Optimizer: paged AdamW 8-bit
- Gradient checkpointing: enabled
- Seed: 42
The run was recovered once from the model-only 30K checkpoint after external
termination. The recovery contract is included and records that model and
trainer step were restored while unavailable optimizer momentum was reset.
Checkpoint layout
The repository root is the original checkpoint-40000 layout. It contains
two safetensor shards, model/processor configuration, modality metadata, and
the saved experiment configuration. Optimizer and scheduler tensors are not
included because this is a model-only checkpoint.
Serving and evaluation contract
The checkpoint decodes 30-step action chunks. The parity evaluation recipe
commits 20 steps at 50 Hz and replans at 2.5 Hz with temporal ensembling off.
The bridge core is byte-identical to the previously deployed Piston bridge;
only the isolated Job A profile is added.
The reference Trial 32 closed-loop run completed all 1,025 action boundaries
and all three 50 FPS videos, so transport and camera contracts passed.
Known limitation
The reference run did not lift the Piston. Absolute hand targets also chatter
at replanning boundaries: the low-frequency task trend is present, but new
chunks can overwrite the previous hand target with a large jump. Treat this
checkpoint as an experimental artifact, not a validated successful policy.
The recommended next evaluation adds a hand-only rate limiter derived from
the demonstration velocity envelope while preserving deliberate release and
regrasp transitions.
Reproducibility files
See
recipe/README.md for the file map. SHA256 manifests
are provided for the uploaded artifact and the original source code contract.
