My research goal is to see complex, human-like behavior emerge from unsupervised interaction between groups of learning agents with an applications focus on enabling autonomous vehicles to operate in rich scenarios. Concretely this leads to a lot of questions:
- How can we use RL to design models of human agents?
- How can we ensure that RL-designed agents are human-compatible?
- How can we synthesize environments that push and test the capabilities of our agents?
- What algorithmic advances and software tools are needed to address these questions?
In practice this means working on understanding how to push the state of the art in multi-agent RL algorithms, designing new data-driven simulators, and trying to deploy simulator-designed controllers into real-world systems.