MPC with Sensor-Based Online Cost Adaptation

This is the code base used for the online cost learning project. The code contains the Bilevel optimization implementation that can generate the optimal weights for the kuka reaching task w/o obstacles. It also has the pytorch implementations to train the vision encoder and the QPNet. Finally, it contains code that is used to deploy the MPC algorithm on the real kuka robot.

The code used to train the QPNet and generate data can be found inside the Notebook directory. The code to train the encoder can be found in the vision directory. The demo directory contains the implementation used to deploy the algorithm on the robot. The python directory contains the bilevel optimization problem code along with the various cost functions implemented in python to generate the ground truth reaching motions on the kuka.

Dependencies

1. pytorch
2. dynamic_graph_head - https://github.com/machines-in-motion/dynamic_graph_head/
3. robot_properties_kuka - https://github.com/machines-in-motion/robot_properties_kuka/

Citation

If this code base is used in your research please cite the following paper

@article{meduri2022mpc,
 title={MPC with Sensor-Based Online Cost Adaptation},
 author={Meduri, Avadesh and Zhu, Huaijiang and Jordana, Armand and Righetti, Ludovic},
 journal={arXiv preprint arXiv:2209.09451},
 year={2022}
}

Maintainer

1. Avadesh Meduri
2. Huaijiang Zhu

Copyrights

Copyright(c) 2023 New York University

License

BSD 3-Clause License