Multi-Fidelity Bayesian Optimization for Simulation-Based Autonomous Driving Systems Testing

Summary

Simulation-based testing of autonomous vehicles has been widely studied in both academic and industrial contexts. It relies on high-fidelity simulations that mimic the real world to assess the safety and reliability of autonomous vehicles (or part of the autonomous driving systems under test). But "do we really need such computationally expensive high-fidelity simulations for all possible scenarios?"

Turns out, we can reduce simulation costs while maintaining test effectiveness by dynamically selecting an appropriate simulation fidelity level for a given scenario. More details available in the preprint/podcast.

This paper is just the first step towards leveraging multi-fidelity optimisation in simulation-based AV testing. We already have a few ideas to improve, and are open to more collaborations!