SO101 ACT leader simulation baseline

ACT trained for 2,000 optimizer steps on 37 leader-controlled simulated pick-and-place demonstrations using LeRobot 0.4.4, on one NVIDIA L40S. This is an initial baseline; task success has not yet been evaluated.

  • Task: Put the cube on the tray.
  • Inputs: scene and wrist RGB cameras, each 640 by 480, plus six joint positions.
  • Control rate: 30 Hz. Five arm joints use degrees; the gripper uses 0 to 100.
  • Training batch: 8. Learning rate: 0.00001. Chunk size: 100.
  • The included custom policy.py enables ACT temporal ensembling at 0.01 and clamps the denormalized gripper target to 5 through 50.
  • Keep the complete processor and configuration files with the model weights.
  • Resume weight fine-tuning with lerobot-train --policy.path=CHECKPOINT_DIR.

The policy.py inference settings require simulation and real-arm validation. Eligibility of work prepared before the event briefing has not been confirmed.

ACT method: Tony Z. Zhao, Vikash Kumar, Sergey Levine and Chelsea Finn, Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware (2023). Implementation: Hugging Face LeRobot. Environment: NWaves SO-101 hackathon.

Downloads last month
3
Safetensors
Model size
51.7M params
Tensor type
F32
·
Video Preview
loading

Dataset used to train ppprock11/so101_act_leader_sim_v1