NVIDIA Lays Out Full-Stack Halos Safety System for Scaling Physical AI
NVIDIA Blog · rss · 2026-09-22
NVIDIA argues that scaling physical AI—49 million L3-5 AVs by 2035 and 60 million industrial robots by 2035, per ABI Research and Omdia—demands safety across hardware, software, AI behavior, and the operating environment, not a one-time pre-deployment check.
Four shifts define the new safety model
- Dynamic, shared environments require context-aware safety rather than static barriers
- AI behavior needs its own assurance, with emerging standards like ISO/IEC TS 22440
- Deployment is ongoing; model updates may require re-validation
- Validation at scale requires simulation, synthetic data, and scenario reconstruction
NVIDIA's Halos stack
- AVs: DRIVE AGX Thor safety compute, Hyperion L4 reference platform, ASIL-D certified Halos OS on DriveOS, open Alpamayo reasoning VLA models, and the Halos Safety Evaluation Framework
- Robotics: IGX Thor with a dedicated Functional Safety Island, Halos Core, Holoscan Sensor Bridge, Isaac Lab/Omniverse simulation, and the open source Outside-In Safety Blueprint for facility-level monitoring
Ecosystem & certification: Geely, Isuzu, Nissan (Wayve software), and Einride build L4-ready vehicles on Hyperion; Uber, Grab, and Lyft use it for robotaxi scaling; Agility integrates IGX Thor into Digit 5's safety system. TÜV SÜD certified DriveOS 6.0 to ISO 26262 ASIL D, and ANAB accredited the Halos AI Systems Inspection Lab as an ISO/IEC 17020 inspection body.
More from Embodied
- UBTECH sells 400 U1 Pro humanoids to Yongda Auto in ~$10M deal — CyberRobooo · 2026-11-12
- Baidu World 2026 to showcase Apollo Go robotaxi rides and Apollo World physical AI model — Baidu_Inc · 2026-11-12
- NVIDIA's open-source Alpamayo 2 Edge VLA driving model lands in Autoware — drmapavone · 2026-11-12
- A closer look at Astra's robot gripper design (and a joke at its expense) — _Stocko_ · 2026-11-12
- Robot startup passes first presale milestone with 10+ units claimed — R4nsack · 2026-11-12
- Runway launches Praxis-1, a 27B MoE world action model, with a 73ms Fast variant — ChongZzZhang · 2026-11-12