Zoox Engineers Present on Resilience in Autonomous Vehicle Teleoperation Systems

The presentation by Jan Hu, a product security engineer at Zoox, and Shane Gupta, Zoox’s director, focuses on resilience in autonomous vehicle teleoperation systems. Teleoperation involves a human remotely assisting an AI-driven vehicle when it encounters uncertain scenarios, such as construction zones, by providing waypoints rather than direct control. The team validates safety through rigorous testing in simulated environments, replicating complex real-world scenarios to assess AI-human collaboration. To identify hidden risks, they adapted fuzzing—a software security technique—to autonomously generate over 50,000 mutated teleoperation commands, uncovering rare but high-risk collisions caused by subtle errors or malicious inputs. Key findings included edge-case bugs in prediction modules and perception systems, which were fixed by retraining AI decision logic. The approach extends beyond teleoperation, applying fuzzing to sensor inputs and planning modules to proactively stress-test autonomous systems. The goal is to scale safety validation by actively challenging the system, ensuring resilience against both accidental and intentional disruptions.