adityaGautham/act_rgb_plastic_20260811

ACT — SO-101: pick up the plastic box, place it on the teal cap

ACT policy (default lerobot config, 52M params, ResNet18 backbone) trained with lerobot v0.6.1 on 32 teleop episodes (11,179 frames @ 30 fps) recorded 2026-08-11 on an SO-101 follower.

• Cameras (must match at inference): top 960×600, wrist 640×480, RGB @ 30 fps
• State/action: 6-dim joint positions (degrees)
• Data: merge of the six adityaGautham/rgb_plastic_20260811_* datasets; normalization stats recomputed over the merged set
• Training: batch 8, 100k steps, ~2 h on RTX 5090; final loss 0.052 (L1 0.052)

Run

lelab: use adityaGautham/act_rgb_plastic_20260811@root as the model reference, or CLI:

lerobot-rollout \
  --strategy.type=base \
  --policy.path=adityaGautham/act_rgb_plastic_20260811 \
  --robot.type=so101_follower \
  --robot.port=<your port> --robot.id=<your id> \
  --robot.cameras="{ top: {type: opencv, index_or_path: <i>, width: 960, height: 600, fps: 30}, wrist: {type: opencv, index_or_path: <j>, width: 640, height: 480, fps: 30}}" \
  --task="pick up the plastic box and place it on the teal cap" --duration=30

Requires lerobot ≥ 0.6.0 (external normalization processors). Trained on 32 episodes from a single scene — expect sensitivity to camera pose, object appearance, and lighting.