Unitree CEO Wang Xingxing: Whoever builds the first Generalized Robotics Foundation Model deserves a Nobel Prize. He also said Unitree is working in that direction too, but the probability of succeeding is currently below 50%.

The hard part is the embodiment gap. Text data transfers across LLMs because the interface is shared. Robot data does not transfer cleanly when hands, joints, sensors, payloads, and environments differ. That is why “robot foundation model” is a much harder claim than it sounds.

There are already tons of generalized robotics foundation models out there, depending on how you define generalized. The model can perform 10, 100, 1000… industrial/domestic tasks?

Most robotics “foundation models” today are still highly constrained systems trained in narrow environments. Scaling from 10 tasks to 1000 tasks doesn’t mean true generalization. Most are only trained on billion-level tokens, which is tiny compared to frontier LLMs. Real generalized robotics models would look much closer to Physical AI or AGI: adapting to open-world environments and handling most real-world tasks humans can do.

Could you please send me a full link for this interview

Thanks for sharing

It will probably be years and manny innovations along the way. When is Unitree going public?

@dennisihong Ahem ..

We wrote something about how the early bet on Unitree first took shape in 2017.

Unitree CEO says whoever builds the first generalized robotics foundation model deserves a Nobel Prize

I don't understand how he mathematically derived or statistically estimates that 50% maybe is in the original Chinese interview ...

A true generalized robotics foundation model would probably be one of the biggest breakthroughs in engineering history.

今のロボットが、目の見えない人が訓練によって得た手の動きを再現する、モーション・ライブラリー的な制御であるならば、とてもじゃないけどランダムに発生するイベント、対象物に対応出来ない。危な過ぎて、家庭に1台なんて夢のまた夢。