act_piper_pour_rice
robot control model for pouring rice
2 min readupdated
This model is a control policy for robotics, trained using the ACT method (Action Chunking with Transformers — predicting short action sequences instead of individual steps). The solution is built on the LeRobot framework and trained on teleoperation demonstration data to perform a household task. The model takes robot state data and images from three cameras (wrist, chest, and top) as input and outputs control actions. It is suitable for researchers and developers working with physical robotic manipulators and imitation learning.
Key features
- ACT architecture: Transformer-based imitation learning for predicting action chunks.
- Multimodal input: Using data from three different high-resolution cameras.
- Training dataset: Trained on a specialized dataset of 30 episodes (48,715 frames).
- LeRobot integration: Full compatibility with the LeRobot ecosystem for deployment and fine-tuning.
What you can do
- Run on a robot: Apply the ready-made policy to perform the rice-pouring task on a compatible robotic manipulator using LeRobot command-line tools.
- Fine-tune the model: Use the configuration and dataset to train your own policy using LeRobot utilities.
Example
Running the ready-made policy on a real robot using a utility from LeRobot:
lerobot-rollout \
--strategy.type=base \
--robot.type= \
--robot.port= \
--robot.cameras="{ : {type: opencv, index_or_path: , width: 640, height: 480, fps: 30}}" \
--policy.path=advanced-control/act_piper_pour_rice \
--task="task" \
--duration=60Read more
DetailsLicense: apache-2.0
Optimizer: adamw (learning rate: 1e-05, batch size: 8)
Training steps: 100,000 at 15 FPS
Input dimensions: state (7,), images (3, 360, 640)
Output dimensions: actions (7,)










Comments