RoboRacer is an international community performing autonomous systems research through building and racing of 1:10 scale vehicles. The common goal is to race around tracks as fast as possible – usually first in simulation and then in the real world. This provides a platform for addressing many research and systems engineering questions and solving both theoretical and practical problems. Since the dynamics and sensor models are often imperfect and potentially even unavailable in these settings, learning from data is part of most modern solutions to autonomous racing. This is further supported by the large amounts of data and fast computers available with modern technology.
One popular approach is to use reinforcement learning, where the idea is to learn parts (or all) of the autonomous driving pipeline solely from data. For example, the planner, the estimator, or even both could be designed in a model-based way – effectively moving these subsystems to the RL agent’s environment. Similarly, it is possible to treat the low-level controller externally, where the RL agent is then used only as a high-level planner. Which one of these hybrid paradigms achieves the best performance strongly depends on the context and is still an open research question. The aim of this project is to contrast and evaluate these different configurations for the task of overtaking in multi-vehicle racing on a physical track.