act_box_all
robotic arm control model for sorting objects
1 min readupdated
Trained with the LeRobot library for the so_follower robotic arm, this policy (a trained model for robot control) handles moving red cubes and boxes along a predefined scenario. Using teleoperation data, it predicts short action sequences (action chunks) by combining visual camera feeds with the manipulator's current state. It is aimed at robotics developers and researchers working within the LeRobot ecosystem.
Key features
- Uses two cameras (overhead and wrist-mounted) at 480x640 resolution.
- Controls 6 degrees of freedom on the robotic arm.
- Trained on a dataset of 105 episodes (around 39,000 frames).
- Built on the Action Chunking with Transformers (ACT) algorithm.
What you can do
- Run the ready-to-use policy to move cubes on your so_follower robot.
- Explore the training dataset using the LeRobot visualizer.
- Fine-tune the model on your own demonstrations for new manipulation tasks.
Example
Running the pre-trained policy on the robot:
lerobot-rollout \
--strategy.type=base \
--robot.type=so_follower \
--robot.port= \
--policy.path=charlilou/act_box_all \
--task="Put the red cube 3" \
--duration=60Read more
DetailsOptimizer: adamw
Learning rate: 1e-05
Batch size: 16
Training steps: 19000
Frame rate: 30 FPS










Comments