UC Berkeley

As robotic systems become more advanced and their applications more complex, it is insufficient to consider a robot's behavior in isolation. Robots are increasingly expected to operate in environments populated with other intelligent agents. In order to perform their tasks well while obeying constraints, these robotic systems must be endowed with the ability to predict the behavior of the other agents in the environment in addition to planning their own actions over a time horizon. This is especially important in scenarios where agents do not communicate with each other and may behave in an adversarial manner. In this thesis, we investigate methods for prediction and planning for multi-agent systems over a variety of reward structures and information structures. Namely, we examine approaches which tackle the problem in cooperative, non-cooperative, and competitive scenarios where agents engage in partial or no communication about their intentions and future plans. These approaches are formulated through a combination of model-based optimal control and data-driven learning techniques, where we use data in a principled manner to construct or augment the objective and constraint functions of optimal control problems. This allows us to incorporate the rich and expressive behavior stemming from learned models in a transparent manner. We place a particular emphasis on the problem of autonomous racing, which is highly illustrative of competitive multi-agent settings with no agent communication and where both prediction and planning are paramount to achieving good performance while maintaining safety in the presence of adversarial agents. The presented approaches are evaluated in simulation and hardware experiments of vehicle navigation and racing tasks.