Leveraging non-smooth multibody dynamics and deep reinforcement learning to infer control policies for autonomous robots and vehicles

This work is the result of synergies between Multi Body simulation and Deep Reinforcement techniques for continuous control. We developed a Python module for physics simulation that wraps the Project Chrono library and we leveraged it to build a set of Reinforcement Learning environments. We implemented a state of the art Deep Reinforcement Learning algorithm capable of dealing with heterogeneous sets of input tensors and used it to solve the environments we built. The tasks solved include robotic control and autonomous driving with sensor fusion for navigation in unknown environment.