We are a physical intelligence company building recursive co-improvement systems where world models and embodied agents evolve together.
Powered by efficiency-native AI models and infrastructure, we drive the collaborative improvement of environments and embodied agents, bringing intelligence from virtual worlds into the physical world.
World models and embodied agents must improve together. Isolated progress in either hits a wall; recursive feedback between them unlocks generalization.
Model-Infra co-design is not optional. By optimizing training and inference together, we make recursive physical intelligence economically viable.
We scale learning in virtual worlds, verify what matters in reality, and close the loop so both worlds continuously improve each other.
We publish fundamental research and open-source tools while building enterprise-grade products that bring physical intelligence to industry.
Our team combines deep expertise in generative world models, embodied agents, and large-scale AI infrastructure. This unified perspective lets us co-optimize algorithms and systems in ways that isolated teams cannot.
Training Cost Reduction
Through co-designed distributed RL and rollout systems
Inference Cost Reduction
Via streaming inference, compression, and hardware-aware optimization
Recursive Co-Improvement
The first systematic loop for world-model and agent co-evolution