Sergey Levine on robotics: hardware is good enough, the real gap is decision-making and data

机器之心 · wechat · 2026-10-05

Figure's Index data platform hit 16M+ human task videos from 108 countries in four months; Figure paid $15M and plans $1B+ in data/compute over the next year.
Physical Intelligence co-founder Sergey Levine argues:
Hardware is now good enough (Boston Dynamics was about control, not decisions); the frontier is the decision loop — AI must understand the world beyond the robot's body.
Priors matter: Google's "arm farm" showed tabula-rasa learning plateaus at grasping; robots need priors from observation and experience, like humans assembling IKEA furniture who've seen furniture before.
Counterintuitive data ordering: ground models in real embodied experience first, then they absorb simulation and internet video better — opposite of the popular YouTube-first route.
Reliability is existential: unlike LLMs, robots must complete tasks autonomously to be useful; RL is key to going from 95% to 100% success, still unsolved.
Ecosystem and China: China shipped 40k+ humanoids (97% of global) in H1, but the whole field has only 100k hours of embodied data; no single lab can build a foundation model alone — data must be shared.
Robotics hasn't entered its stable scaling era yet; the industry must first find what's genuinely worth scaling, with Wang Xingxing estimating a "ChatGPT moment" is 2-5+ years away.

More from Embodied