This assignment investigates the following main research question:
How to learn interpretable physical interaction policies using RL?
This will include:
- Design and implement the simulation learning environment in Isaac Lab.
- Investigate the proper MDP design (observations, actions, rewards, domain randomisation) for learning interpretable physical-interaction control policies.
- Investigate the proper feature library (e.g. polynomials, trigonometrics, etc) for learning a controller for this task.
- Investigate the difference between direct-control policies and policies that tune a low-level controller.
- Compare the performance with benchmark controllers such as classical controllers and learned policies (non-interpretable).
- Experimental validation of the learned policies.