Learning-based Bayesian Inference for Testing of Autonomous Systems

Project Link

Sampling based methods such as MCMC are often used for exploration of failures in autonomous systems. However, these methods suffer from inefficiency in high dimensional exploration, and complex search spaces. In this work, we investigate two principled techniques for improving the efficiency of sampling process for falsification. First, we use a Neural Projection Operator (NPO) and VAE for constrained sampling within a well-defined target region, and high dimensional exploration respectively. Second, we use second-order gradient based sampling, callede Langevin algorithm, to improve the speed of convergence, and prevent entrapment in local minima, a common pifall of gradient-based optimization.

We apply our overall approach on several autonomous racing and trajectory tracking tasks (F1-Tenth and Autorally) with different controllers and dynamic systems, and observe substantial improvement in rate of failure discovery across all tasks. This work was presented in ICRA 2025 and published in RA-L 2024. The project website can be found here.

Recommended citation: A. Parashar, J. Yin, C. Dawson, P. Tsiotras, C. Fan (2024). "Learning-based Bayesian Inference for Testing of Autonomous Systems." IEEE Robotics and Automation Letters.
Download Paper