wbell7/starkhacks_act_cell_step20k

StarkHacks 2026 — ACT policy, step20k

Task: Pick the orange cell and place it into the empty slot of the green module.

Rank: deeper plateau sibling of 15K. Just-past-plateau checkpoint. Test whether 2-3 more epochs of post-plateau training helped or hurt vs step15k. Training loss identical; parameter positions differ.

Trained on wbell7/starkhacks_cell_pickplace (58 episodes, 29K frames, cameras: top + wrist, 30 fps, 640×480).

Run: https://wandb.ai/wrbell7/starkhacks/runs/nbyv34do

metric	value
step	20000
epoch	21.94
train loss	0.118
batch size	32, bf16
policy	ACT, ResNet-18, 52M params

Usage

lerobot-record \
  --robot.type=so101_follower --robot.port=/dev/so101_follower --robot.id=my_follower_arm \
  --robot.cameras="{top: {type: opencv, index_or_path: /dev/cam_ugreen, width: 640, height: 480, fps: 30}, wrist: {type: opencv, index_or_path: /dev/cam_arc, width: 640, height: 480, fps: 30}}" \
  --teleop.type=so101_leader --teleop.port=/dev/so101_leader --teleop.id=my_leader_arm \
  --dataset.repo_id=local/eval_step20k \
  --dataset.single_task="Pick the orange cell and place it into the empty slot of the green module." \
  --dataset.num_episodes=5 --dataset.episode_time_s=30 --dataset.reset_time_s=20 \
  --policy.path=wbell7/starkhacks_act_cell_step20k \
  --dataset.push_to_hub=false --display_data=false