Physical intelligence · World models
Cirquar builds generalizable world models for embodied intelligence, with home service as the core deployment. We work to break the generalization bottleneck of world models in open physical environments — so robots can carry out long-horizon tasks in real homes, safely and flexibly.
01
Real-world data is scarce by nature. A world model must amplify limited episodes into transferable regularities: after one experience, more than a trajectory is remembered; after one law is mastered, unseen situations can be imagined.
02
An embodied system always faces a partially observable world. At every moment the machine should answer: what state the world might be in, what might happen next, and how confident it is in those judgments.
03
Belief must enter action. Not by consuming reality through trial and error, nor by demanding that reality run under preset conditions — but by acting in ways fitted to an open environment.
World models are the only path to general embodied intelligence.Finite data cannot enumerate the infinite variation of reality.
Imitation can acquire motions quickly. Data can cover common scenes. Yet the open world will always produce states beyond the training set. General embodied intelligence must grasp the laws that generate those variations.
Today we take nearly a decade of work from the laboratory into the physical world — to bring embodied intelligence into everyday life.
We are looking for early members in world models, reinforcement learning, and embodied systems — people willing to let machines form beliefs about the world, in open physical environments.