The advent of state-of-the-art machine learning models and complex human-robot interaction systems has been accompanied by an increasing need for the efficient generation of diverse and challenging scenarios to test these systems.

In this talk, I will formalize the problem of algorithmic scenario generation and propose a general framework for searching, generating, and evaluating simulated scenarios. I will first discuss our fundamental advances in quality diversity optimization algorithms that search the continuous, multi-dimensional scenario space. I will then show how integrating quality diversity algorithms with generative models allows for the generation of realistic scenarios. Instead of performing expensive evaluations for every single generated scenario in a robotic simulator, I will discuss combining the scenario search with the self-supervised learning of surrogate models that predict human-robot interaction outcomes. Finally, I will introduce the notion of 'soft archives' for registering the generated scenarios, which significantly improves performance in hard-to-optimize domains.

While the talk will focus on scenario generation, the proposed framework is general and can be applied to a wide range of applications where diverse datasets are desirable. I will conclude the talk by discussing applications in searching for diverse faces, robot locomotion policies, and warehouse layouts.