Physical AI researcher: even 99% generalization cuts per-task cost by only 50%
m_wulfmeier · x · 2026-09-06
At the IWIALS workshop, NoMagic AI's Wulfmeier argued that while generality has advanced, physical AI must now demonstrate real-world results to justify funding.
Using Chinchilla scaling as a (pessimistic) lens, he noted that even generalizing a task to 99% only reduces per-task budget by 50% — optimistic versions still leave 30% — so new recipes and data sources are needed.
The team is also hiring research scientists and engineers.
More from Embodied
- Suzhou's Zeroth Robotics launches sub-$4,000 humanoid robot to open its platform to developers — CyberRobooo · 2026-09-06
- Autonomous OS open-sources on GitHub: an 'Android for robots' that runs agentic reasoning onboard — dee_hw · 2026-09-06
- Why I just had to buy that Robot Duck (even if it gathers dust) — erikstarck · 2026-09-06
- LG and NC AI Team on Autonomous Humanoid; Gemini Spark Privacy Concerns — emmanuelvivier · 2026-09-06
- Meta's Muse Voice Transcribe: 80ms streaming transcription, cheapest and most accurate — The Decoder · 2026-09-06
- OpenAI hardware reveal rumored for DevDay, with launch planned for early 2027 — kimmonismus · 2026-09-06