wbell7/starkhacks_act_cell_step15k

StarkHacks 2026 — ACT policy, step15k

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

Rank: #2 (tighter fit — just before plateau). Lower loss than step 10K but approaching the overfitting inflection. Try this if step 10K misses task nuance.

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

Training run: https://wandb.ai/wrbell7/starkhacks/runs/nbyv34do

Metrics at this checkpoint

metric	value
step	15000
epoch	16.46
train loss	0.122
batch size	32
optimizer	AdamW, lr=1e-5, wd=1e-4
mixed precision	bf16
hardware	AMD MI300X VF (ROCm 7.2, torch 2.7.1+rocm6.3)
policy	ACT, ResNet-18 backbone, 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}}" \
  --dataset.repo_id=local/eval_step15k \
  --dataset.single_task="Pick the orange cell and place it into the empty slot of the green module." \
  --dataset.num_episodes=3 \
  --policy.path=wbell7/starkhacks_act_cell_step15k \
  --dataset.push_to_hub=false \
  --display_data=false