act_open_box
a robotic policy for opening box lids
2 min read
This model is a trained control policy for robots (specifically for the so_follower configuration) designed to tackle the task of opening a box lid. It is built on the ACT approach (Action Chunking with Transformers—an imitation learning method that predicts short action sequences instead of single steps) and the LeRobot library. Created using a dataset of teleoperation demonstrations, the model outputs control actions based on the robot's state and images from two cameras (wrist-mounted and overhead). It is useful for researchers and developers working with real-world robotics based on the LeRobot library.
Key features
- Trained using the ACT method based on Transformers to predict action sequences
- Uses visual observations from two cameras: wrist and top
- Compatible with the so_follower robotic configuration via the LeRobot framework
- Trained on a specialized dataset of open_box demonstrations
What you can do
- Run the ready-to-use policy on a compatible robot to automatically perform the task of opening a box lid
- Use the configuration and dataset for fine-tuning your own robotic models within the LeRobot ecosystem
Example
lerobot-rollout \
--strategy.type=base \
--robot.type=so_follower \
--robot.port= \
--robot.cameras="{ : {type: opencv, index_or_path: , width: 640, height: 480, fps: 30}, : {type: opencv, index_or_path: , width: 640, height: 480, fps: 30}}" \
--policy.path=Sergio23g/act_open_box \
--task="Open the box lid" \
--duration=60Read more
DetailsDataset: Sergio23g/open_box (10 episodes, 4116 frames at 30 FPS)
Training parameters: 10,000 steps, batch size 8, adamw optimizer, learning rate 1e-05
Inputs: state (dimension 6) and images from two cameras (3, 480, 640)
Outputs: control actions (dimension 6)










Comments