Model-predictive control repeatedly evaluates possible actions before selecting what the robot should do next. We explored using Isaac Gym's GPU-parallel simulation as the dynamics model for a sampling-based controller.
Approach
I created the controller together with Corrado Pezzato and Elia Trevisan. Isaac Gym made it possible to run many candidate robot trajectories in parallel, score their outcomes and use those samples to update the control sequence.
Research and open source
The work received third-best paper at a DeepMind-hosted workshop in London during ICRA 2023. As the main developer and maintainer of the open-source software, I developed the implementation beyond the paper and supported a project that went on to receive more than 150 stars on GitHub.
What shaped my work
This project combined controls research, GPU simulation and software maintenance. A useful research implementation needs not only a strong method, but a codebase that others can understand, run and build upon.