Shengshu's Zhu Jun bets on video for robots
Shengshu's Zhu Jun says embodied AI should learn general world knowledge from large amounts of video before it is aligned on robot data, which could reduce reliance on costly teleoperated trajectories.
Shengshu's Zhu Jun used a CCTV interview at WAIC 2026 to lay out how the company trains embodied AI. The interview announced no new model, benchmark result or funding.
Zhu's starting point is the cost of robot data. Conventional vision-language-action (VLA) models are trained on trajectories collected through teleoperation and then fitted to specific actions through imitation learning. According to the interview, this works well in constrained settings, but generalization often drops sharply when objects, backgrounds or robot embodiments change.
Shengshu reverses the usual order. Its models first learn general world knowledge from video, and robot data is used afterward to align specific skills and embodiments. Zhu calls the data mix a "data pyramid" with four layers:
- internet-scale general video
- first-person video of human actions
- simulation-synthesized data
- real robot trajectories
General video covers people, objects, environments and motion. First-person footage is closer to how humans manipulate objects. Simulation and robot data supply embodiment state, control commands and precise action feedback.
Internet video normally has no action labels, so it is hard to use directly. Shengshu handles this with a generative objective: the model learns by predicting and reconstructing pixels. Zhu compares this to next-token prediction in language models. He argues that pixel-level prediction is harder but forces a model to handle objects, space, time, actions and state changes all at once.
Zhu linked this approach to U-ViT, an architecture his team proposed and open-sourced in 2022 that brought Transformer attention into diffusion models.
The interview did not say how much robot data the approach saves, and it gave no evaluation results.