What is it ?

Code for Differential Value Programming (DVP) implementation and example on the KUKA LBR IIWA 14 manipulator.

Dependencies

• PyTorch >= v1.8
• Crocoddyl >= v1.8.1
• Python 3.6
• Pinocchio >= v2.6.3

How to use it ?

Install PyTorch, Crocoddyl and Pinocchio first if you don't have them already installed, then

https://github.com/amitparag/Kuka-arm-DvP.git

Then cd exp_arm/ and

• First generate a test set by running python datagen.py
• Then launch the training using python main.py

The trained NN will be saved in results/trained_models and figures of the training sets generated throughout iterations will be saved in results/figures.

The directory config/robot_properties_kuka contains URDF and meshes information of the robot, and config/ocp_params contains sets of parameters describing the OCP for Crocoddyl. The OCP is setup in utils/ocp_utils.py

Acknowledgements

You can find more information on this algorithm in this paper :

Parag, A., Kleff, S., Saci, L., Mansard, N., & Stasse, O. Value learning from trajectory optimization and Sobolev descent : A step toward reinforcement learning with superlinear convergence properties, International Conference on Robotics and Automation (ICRA) 2022 [accepted]