Mowing Down Simulated Elephants Could Help Self-Driving Cars Prepare For the Chaos of Real Life Streets

Researchers introduced Fail2Drive, a new CARLA-based benchmark for testing self-driving cars in unusual and random hazard scenarios.

The benchmark adds rare, bizarre road obstacles such as an elephant, firetruck, and playground slide to probe edge cases.

Testing showed model success rates dropped by 22.8%, revealing weaker robustness than standard benchmarks suggest.

The results highlight a major safety gap: autonomous vehicles can look strong in familiar tests but fail on out-of-distribution conditions.