3-month-old Simate tops RoboDojo benchmark with AI-driven recursive self-improvement for robotics
智东西 · wechat · 2026-09-24
Physical intelligence startup Simate, founded just three months ago, has topped the RoboDojo robotics benchmark with a general physical fast-system model developed partly by its AutoResearch AI research system, and raised hundreds of millions of RMB.
- Team: founder Zhang Ying from a leading autonomous driving company; HKUST assistant professor Zhang Fangneng (world models); young scientist Ji Ma Zeyu, formerly of ARI (acquired by Meta)
- Real-robot demos show long-horizon tasks (like brewing tea), memory and high-precision manipulation, using 4D physical perception and memory mechanisms
- The company proposes PhysicalRSI (recursive self-improvement for physical intelligence), splitting research tasks into weak/medium/strong tiers, with AI agents handling model modification, experiments and evaluation
- Tech stack includes the pluggable SiPAI model framework (world models, VLA, VLM) plus a closed loop of world-model simulation and real-robot validation
- AutoResearch is in trial use by researchers at Tsinghua, MIT and HKUST
The approach mirrors Anthropic's disclosure that Claude "leads" 26% of AI R&D under human supervision, but applies it to robot intelligence research.
More from AGI Musings
- Dev Who Weathered 3D Accelerators and Win95 Says AI Hits Harder — gandamu_ml · 2026-09-24
- UK Officials Privately Believe AI Companies Overstate Risks for Financial Gain — connoraxiotes · 2026-09-24
- Commenter: 90% of Silicon Valley Is in Mass Psychosis Over AI, Backed by Trillions — mjdramstead · 2026-09-24
- LWiAI Podcast #257: GPT-6 Astra, AI Extinction Warnings, and a Wave of Security Incidents — Last Week in AI · 2026-09-24
- Dance duo Jess and Morgs stage 'Deepfake', interrogating truth in the AI era — nordicinst · 2026-09-24
- Australian Tutoring Company Shuts Down, Tells Parents to 'Use AI Instead' — Old-Competition3596 · 2026-09-24