Waymo explains at Hot Chips why camera-only autonomy plateaus at 99.9%
4310sy · x · 2026-11-09
At Hot Chips, Waymo Compute Lead Daniel Rosenband explained the motivation behind the company's custom Sensor Fusion Processor in a talk titled "Compute in Motion."
- Camera-only limits: pure-vision approaches reach decent results quickly but plateau; a classic example is a red balloon floating in front of a traffic light being misread as a red signal — something even Gemini still struggles to disambiguate.
- The last 99.9%: Rosenband argued today's Physical AI (including Waymo) is only the first 1%; productizing means stacking nines (99% → 99.9% → 99.99%), and "doing the same thing" won't get you there — hence the pivot to multi-sensor fusion.
- Track record: by 2020 Waymo had driven 200M+ miles (320M km), completed 20M+ paid rides across 15 cities, and is now on 6th-generation hardware, with a fully driverless service planned in Japan (with Nihon Kotsu and GO) for 2027.
The reposter adds that only Waymo, already climbing that plateau, is qualified to make this judgment.
More from Embodied
- 'The future is here, just unevenly distributed': humanoid robot spotted at a conference — technollama · 2026-11-09
- Developer builds a physical prompting controller out of OpenAI DevDay merch — paw_lean · 2026-11-09
- Robot installations hit 603k in 2025 as China holds 59%, US second — ingliguori · 2026-11-09
- Zipline raises $1.5B at $21.5B valuation to build nationwide drone delivery network — emmanuelvivier · 2026-11-09
- Touch-R1 uses RL for cross-sensor tactile reasoning, beating GPT-4o by 24.7 points — 机器之心 · 2026-11-09
- CoRL 2026 spotlight paper explains why VLAs fail at contact-rich tasks and how to fix them — StanfordAILab · 2026-11-09