NVIDIA is making the case that robotaxi safety has to be designed into the full system, not bolted on after deployment. Its HALOS approach focuses on safety across the autonomous-vehicle stack as robotaxi services move from pilots toward broader commercial use.

That matters because robotaxis combine AI perception, planning, hardware, simulation, and operational controls in public environments where failure is highly visible. A single model improvement is not enough.

The broader lesson for AI systems is that safety becomes an architecture question when models are tied to real-world actions.