Six Months In, Li Hongyang's Team Shows a Robot That Fetches Its Own Step Stool

Six months after its founding, Li Hongyang's team has for the first time systematically shown that its Whole-Body Intelligence (WBI) approach works on a real robot. The demo is simple in outline: a toy sits too high, and the robot's arm can't reach it. The robot finds a stand, carries it over, climbs onto it and takes the toy. The publication came out on 10 October 2026, and coverage of it spread through the outlet "机器之心" (Jiqizhixin).

What exactly was shown

The core of the video is a sequence of three actions, each with its own physics:

• carrying the stand: the robot holds an object, and its center of mass shifts;
• climbing onto the stand: ground contact and support area change, the platform is higher, and stability is lower;
• taking the toy at height: the arm works in a new workspace, while the robot must keep its balance on a narrow footing.

According to the announcement, this is the first systematic publication of results on a real robot. Before this, the team had only described the WBI technology path in a technical blog. The available excerpt gives no figures for episode duration, success rate, platform type or degrees of freedom. For now, only the scenario can be evaluated.

Why the stand is harder than it looks

Most manipulation demos rely on a simplification: the body stands still while the arm works. That won't work here. When the robot climbs onto the stand, it effectively changes its own morphology. Height, support points and permissible joint forces all change. The controller has to account for this in advance: the stand must be placed so that the robot can later step onto it and reach the toy.

This is the logic behind WBI. If motion planning for the arms, legs and torso is split across separate modules, the task falls apart at the seams: the manipulator doesn't know the support is about to change. A single model or a single control loop should, in theory, remove this problem. In theory, because the video doesn't show how the system is actually built: an end-to-end policy, a hierarchy, or classical whole-body control with learned add-ons.

The choice of scenario itself signals ambition. Using a tool from the environment, even one this simple, requires an understanding of the scene's geometry and of the consequences of the robot's own actions. That is closer to reasoning about physics than to replaying a memorized trajectory. But a single successful episode may also be the product of careful staging.

What to watch next

Telling a scientific result from a polished video takes answers to several practical questions:

• Repeatability. How many attempts out of a hundred succeed, and how does the robot behave when it fails?
• Generalization. What happens if the stand's height or weight, the floor surface or the toy's position changes?
• Autonomy. Who sets the goal: does a human give the command, or does the robot decide on its own that it can't manage without the stand?
• Platform. Is it the team's own body or an off-the-shelf humanoid, and what does a fall cost?
• Openness. Will a paper, code or benchmark be released that lets other labs verify the result?

For the industry, the signal lies elsewhere. The humanoid market is shifting from single-arm demos to tasks in which the whole body takes part in the solution. For factories and warehouses, these are the skills that determine whether a robot can work in a non-ideal environment. Li Hongyang's team now has to show that its approach holds up across dozens of variations, not just one scenario. For now, this is a claim backed by a working prototype, and new publications with metrics will be the real test.