zjumty/so101-act-train007a

SO-101 ACT policy — TRAIN-007A

An ACT (Action Chunking with Transformers) policy for the SO-101 arm, trained in MuJoCo on a pick-and-place task: pick an orange cup on the table and place it on a fixed target.

Inputs and outputs

• Observations: two RGB cameras, side (world-fixed) and wrist, both 640×480, plus a 6-D joint state.
• Actions: 6-D absolute joint targets at 10 Hz. The chunk size is 50 steps (5 s).
• Execution: run it with --execution-prefix 50, which executes each predicted chunk open-loop in full. That is the default ACT execution.

Training

• Code: LeRobot 0.6.1 ACT, through the project's training adapter.
• Config: training-config.yaml (training-side-wrist-007a-bf16), source commit 5e287317f.
  • chunk 50; dwell compression with idle_threshold_rad 0.001;
  • learning rate 1e-4 (backbone 1e-5);
  • batch 16; 60000 steps; bf16 autocast.
• Data: 103 Teacher (scripted MoveIt expert) demonstrations, rig side-wrist. The dataset repo is zjumty/so101-act-side-wrist-joint103. The split is 83 / 10 / 10 episodes (train / validation / offline test).
• Final validation L1: 0.0291. Offline-test first-action MAE: 0.00647 rad.

Closed-loop results (MuJoCo, prefix 50)

Evaluation	DONE
4 standard points × 5 trials	20/20
rollout_validation, 28 unseen positions	27/28
rollout_test, 20 unseen positions, opened once	20/20
Extrapolation, 12 positions just outside the training region	9/12

Known limitation (R0 sensitivity probe): the policy is brittle to appearance changes. A checker table texture gives 15/28, a blue cup 18/28, and a 1° side-camera rotation 17/28. It is robust to light intensity and direction, to 5 mm camera shifts and to ±3% fovy. All of these are evaluation-only scene variants.

Files

File	sha256
policy.safetensors	ea6e03362d1acfd86004f5e0b8a192aac4d6d35f0b0d2ac74149bdee9d1ff447
bundle.json (normalisation, action contract, config and dataset hashes)	f4056ece15ee4f7c5f6f69f8b4701a510f7eb25d4c7b2890e8dd8d6bb5b0101f
training-config.yaml	91797eea7481939e9e17d5ce959b5aa96c406f0a2a7a66e9eb891b3bfa19982b
metrics.jsonl, training-log.json	training metrics and log

The policy is loaded with the project's build_policy_from_bundle (repo ros-moveit-demo, package so101_demo_py). Experiment ledger references: CP-3776 (training), CP-3777 (parallel rollout and held-out results), CP-3780 (R0).