act_pick_and_place_hair_roll_final_chunk100
robotics control policy
1 min readupdated
This model uses ACT (Action Chunking with Transformers, an imitation learning method that predicts short action sequences instead of single steps) to control a robotic arm. Trained on teleoperation demonstrations for a pick-and-place task (pick_and_place_hair_roll_final), it outputs robot control actions from dual-camera feeds and joint states. Ideal for researchers and engineers working on physical robotics with the LeRobot library.
Key features
- Built on the ACT architecture and trained with the LeRobot library.
- Processes two camera streams (top and wrist) alongside robot state data.
- Designed for the so_follower robot setup.
- Trained on a 60-episode dataset (43,285 frames at 30 FPS).
What you can do
- Deploy the pre-trained control policy on a compatible so_follower robot to automate the task.
- Fine-tune the model on your own demonstrations using LeRobot tools.
Example
lerobot-rollout \
--strategy.type=base \
--robot.type=so_follower \
--robot.port= \
--policy.path=knight7024/act_pick_and_place_hair_roll_final_chunk100 \
--task="pick_and_place_hair_roll_final" \
--duration=60Read more
DetailsLicense: apache-2.0
Parameters: 0.05B
Optimizer: adamw (learning rate: 1e-05, training steps: 100000)
Input format: observation.state (6,), observation.images.top (3, 240, 320), observation.images.wrist (3, 240, 320)










Comments