Qianjue, an embodied AI company with roots at Tsinghua University, has launched Polibrain, a predictive world model for robots. The company unveiled it at the first Physical AI and World Model Conference, which opened in Nanjing on October 20, with founder Gao Haichuan presenting the model in the opening keynote.
Polibrain does not try to generate every future video frame. Instead, it learns relationships between environment states, history and action outcomes, and predicts how task-relevant states will change in a latent representation space. For a robot picking up litter, the company argues, the color of the ground matters less than where the object is, how the arm relates to it and how a grasp will change its state.
The model links perception, state understanding, action prediction, decision-making, execution and feedback into a closed loop. When results diverge from predictions, for example after a failed grasp, it re-evaluates the task from the new state and adjusts its next steps. Qianjue says key-feature decoding lets it track gaps between expected and actual states to help locate faults and recover.
In a demo video provided by the company, a quadruped robot carrying an arm and a collection basket searches for litter outdoors, approaches it, plans the arm motion and places items in the basket. Qianjue says Polibrain is being adapted to quadruped, humanoid, wheeled, single-arm and dual-arm robots for tasks such as companion service, food service, tidying and fine manipulation. The company acknowledges that its reliability in complex scenes still needs fuller real-world testing.
Qianjue positions itself as a robot-brain supplier rather than a hardware maker. Its loosely coupled architecture separates a shared representation of the environment and task from each machine's own control, so different robots, or even connected appliances, can carry out the same intended change in their own way. The company hopes this will also let data from different robot forms feed one model.
The work builds on research Qianjue published with a Tsinghua team at ICML 2025 on when neural networks learn world models, along with newer work on an Effective Degree measure of model complexity and a counterfactual method for deciding which past observations to keep.
SEE ALSO: Peking University and PrimeBot's HIL-UMI · SenseTime Open-Sources SenseNova-RoboRSI