Qianjue unveils Polibrain, a robot world model that predicts before it acts

Qianjue, a Tsinghua-linked robotics company, has introduced Polibrain, a predictive world model built for machines rather than for generating convincing video. The goal is for a robot to reason about what an action will change before it moves, the way a traveller might draw a map and plan a route instead of trying every road.

That puts it at odds with the dominant VLA approach, which maps perception more directly to action. Polibrain learns relationships between environment state, history and action outcomes, then predicts state changes in a task-relevant latent space. For a quadruped robot collecting litter outdoors, the colour of the ground and the texture of the background matter little; the position of the litter, the arm's relation to it and how a grasp changes its state are what count.

The system closes a loop across perception, state understanding, action prediction, task decision, execution and feedback. When execution diverges from prediction, the model re-evaluates the task from the new state and adjusts its next moves; decoding key features also lets it trace the gap between expected and actual outcomes for fault localisation and recovery.

To let different bodies share one brain, Qianjue separates a shared representation of environment and task from each robot's actuation. Understanding can be common; execution is left to the humanoid, wheeled, single-arm or dual-arm platform. The company says it is adapting the model across those form factors for companion services, food service, tidying and fine manipulation.

The work rests on research with Tsinghua University. In 2025 the team published "When Do Neural Networks Learn World Models?" at ICML, asking when a network can recover hidden world structure from observations. LeCun and co-authors published "When Does LeJEPA Learn a World Model?" in May 2026 on related identifiability questions; by publication date, Qianjue's paper came roughly a year earlier.

Methodologically, the team uses polynomial representations to stop models memorising background texture and lighting, and a "causal memory" that decides whether a past observation still affects future decisions. It also proposes Effective Degree, which fits orthogonal polynomials along interpolation paths between real data points and weights components by order, giving a computable measure of functional complexity.

The caveat is that this is a technical foundation, not a validated product. The article itself notes that reliability in complex scenes needs fuller real-world testing, and there are no public quantitative results for cross-embodiment transfer.

World models are becoming a crowded frontier: LeCun keeps betting on predictive approaches, Fei-Fei Li's World Labs is chasing interactive 3D worlds, and Qianjue is pitching physical regularity plus reuse across bodies. What to watch: task success rates in real deployments, performance decay when the model moves to a new robot, and whether the company publishes reproducible evaluations.