Root-cause failure analysis and repair for autonomous systems

Recent years have seen large numbers of learning-enabled autonomous systems, especially autonomous vehicles and drones, deployed in the real world. Unfortunately, these deployments have been accompanied by a corresponding increase in safety-critical accidents involving these systems. As designers of autonomous systems, we must be able to predict the ways in which a system is likely to fail and take steps to prevent those failures before deploying our robots in the wild. Existing tools for verification struggle to search over high-dimensional environmental parameters and provide little guidance on how to preemptively mitigate failures once they are discovered. We propose a diffusion-inspired algorithm that uses differentiable physics-based rendering and simulation to efficiently predict failures and propose policy updates to mitigate the severity of those failures. Relative to state-of-the-art black-box verification and optimization methods, our approach substantially reduces the number of expensive simulation queries needed to both predict and repair safety-critical failure modes, generating safer and more reliable repaired policies.