Autonomous Vehicle Safety: 5 Key Points from NVIDIA's Head of Industry Safety

At a glance

NVIDIA's head of industry safety, Riccardo Mariani, explains how self-driving cars can be made safe. The video covers managing risk across the whole lifecycle, from development to operation and software and AI updates.

Key points

• It explains why counting test miles alone does not show that a self-driving car is safe.
• It says safety needs design-time and runtime guardrails, simulation, validation, and real-road testing together.
• It stresses a full-system approach that aligns hardware, software, and tools across simulation, training, and inference.
• It introduces NVIDIA's safety framework, Halos, and argues that the industry should share processes and models to build standards.

Why it matters

This is useful for readers who care more about how the industry can prove self-driving cars are safe than whether they can be safe at all.

More from this field

• Hyundai's Atlas Robot Swaps Feet for Wheels, Drifts and Jumps 360 Degrees

  Hyundai Motor Group's humanoid robot Atlas was fitted with wheels instead of feet in a test. It drifted across the floor and performed a 360-degree spinning jump.

  DongA Science · Nov 14, 08:00 KST
• NVIDIA Unveils GR00T 2, an Open Robotics Model for Researchers

  This video introduces GR00T 2, an open model from NVIDIA made for robotics research. It shows robots running the model in real time.

  NVIDIA 공식 YouTube · Nov 14, 04:36 KST
• NVIDIA Shows How to Run DLA Inference in Python on Jetson AGX Orin with cuDLA

  NVIDIA explained how to use the Deep Learning Accelerator (DLA, a chip that handles neural network work) on Jetson AGX Orin directly from Python. This lets developers move AI work from the GPU to the DLA, saving power and freeing up GPU capacity.

  NVIDIA Technical blog · Nov 14, 03:57 KST
• NVIDIA Releases SimReady Assets for Robot Pick-and-Place Simulation

  NVIDIA shows how to build a robot pick-and-place simulation using SimReady Central assets on Hugging Face. The robot arm, gripper, and can models come with physics settings already in place, so they can be tested in a virtual space right away.

  Hugging Face blog · Nov 14, 03:00 KST