The first Chinese winner of robotics' 22-year-old early-career prize says embodied AI needs the same end-to-end mind that won self-driving

Li Hongyang walked off the stage at the Sydney International Convention Centre on the morning of July 14 with the RSS 2026 Early Career Spotlight. He is the first Chinese scholar to win the award in the 22 years it has existed.

The Robotics: Science and Systems conference is a small, in-person gathering, with a few hundred attendees and single-track sessions. The Early Career Spotlight sits alongside Best Paper and Test of Time as one of the most prestigious honors the field gives. Four scholars received it this year. Li is the unusual recipient: his most-cited work is in autonomous driving, not robotics. He leads UniAD, a CVPR 2023 best paper on end-to-end driving, and co-authored BEVFormer, a perception architecture that became one of the most cited driving papers of 2022.

That track record is the reason his new argument carries weight. In a long-form interview with Leiphone and a separate essay on the Archon blog, Li says embodied AI is repeating the mistake autonomous driving made around 2019: building the system as a stack of separately trained modules (one for perception, one for planning, one for low-level control) and stitching them together at the end.

In his framing, the field has split the robot in two. The upper body gets a vision-language-action model trained to grasp a cup. The lower body gets a contact-rich locomotion policy trained to walk on gravel. Almost no one is training the system as a single whole-body controller that makes one decision, in one forward pass, in real time, on the actual machine.

He calls the missing layer Whole-Body Intelligence, or WBI, and his archon.tech essay lays out a pretraining path for it. WBI is not a humanoid shape. It is a coordinated full-body controller, with arms, legs, perception, and decision-making running through one model. He has also coined a term for the result: 智能人形, which he renders as "intelligent humanoid." Both are his vocabulary, not industry standard.

The bet is large, and Li himself acknowledges it. In the Leiphone interview, he concedes that most of the field considers full-body coordination, on a real humanoid, at foundation-model scale, the hardest open problem in the space, harder than walking or grasping alone.

The pattern worth watching is the cross-domain move. UniAD won its Best Paper at CVPR 2023 because it did the thing the driving field had refused to do: train perception, prediction, and planning end-to-end, on real data, with a single loss. SenseTime's release and Li's own pages credit him as lead. A separate researcher site (romanization differs across his own domains) lists the same work.

The same people are now launching humanoid foundation-model labs. Archon Robotics, the company linked from the WBI essay, closed a seed round reported by yangtzeer in the same week as the RSS award. The bet is the same bet, retargeted.

The falsifier is concrete. If a modular humanoid stack, with separate grasping, walking, and language modules tied together at inference time, scales first to production, the frame breaks. Two things to watch in the next six months: a real humanoid running a single end-to-end model across arms and legs, and any major humanoid lab hiring a planning lead from end-to-end driving.

The next trigger is dated. The Archon team used RSS week to ship the WBI long-form. The seed round is closed. The first peer-reviewed test of the thesis will come when Archon, or one of its rivals, posts a whole-body policy on a real humanoid and the numbers hold.