SamruddhiThorat07/act_green_pilot

ACT policy: SO-101, green cube → basket

An ACT policy trained with LeRobot for a real SO-101 arm. The task is "Put the green cube in the basket".

• Training data: SamruddhiThorat07/green_pilot, 10 teleoperated demos (5,322 frames at 30 fps). The cube starts at one fixed spot and the basket is in a fixed place.
• Inputs: overhead camera (640×360), wrist camera (640×480) and 6-D joint state
• Output: 6-D joint position actions, in chunks of 100
• Training: 3,000 steps, batch size 8, learning rate 1e-5, ResNet-18 backbone, 15 min 45 s on an RTX 3060
• Result: 10/10 successful rollouts at the training layout

This is a single-layout pilot. It hasn't been tested on new cube positions.

Demo: github.com/SamruddhiThorat07/so101-act-green-cube

Use

lerobot-rollout --strategy.type=episodic \
  --policy.path=SamruddhiThorat07/act_green_pilot \
  --robot.type=so101_follower --robot.port=<your port> \
  --robot.cameras="{ overhead: {...}, wrist: {...} }" \
  --dataset.single_task="Put the green cube in the basket" --device=cuda

The camera names (overhead, wrist) and resolutions must match the ones used in training.