Polibrain (千诀科技): predictive world model for robots

Tsinghua-spun embodied-AI company 千诀科技 (Qiansi Technology) released Polibrain on Oct 20 at the first Physical AI & World Model conference in Nanjing: a predictive world model that learns a 'map' of physical dynamics in latent representation space — predicting only task-relevant state changes instead of full video frames. Key choices: polynomial representation for extracting task-relevant variables, causal memory for history selection, and a loosely-coupled architecture letting different robot morphologies share the same world understanding. Demo: a quadruped with an arm doing autonomous trash pickup outdoors.

Why it matters

What you could build with it

Does it hold up?

Too early to judge — released Oct 20; the writeup itself notes real-world reliability needs more complete measured data, and the model is not publicly downloadable.

Built with Polibrain (千诀科技)

• LeCun押注的世界模型: 千诀科技Polibrain发布 (量子位 via 新浪财经)sina · Long-form writeup of the Nanjing launch: latent-space prediction, polynomial representation, causal memory, cross-morphology reuse and the cross-body data flywheel thesis.

Learn more

deep writeup of the launch (量子位 via 新浪财经) →

First spotted on article: source.

More AI models