Research
Digital and physical worlds are the environments. Evolution is the method: interaction creates experience, reinforcement turns feedback into learning, and composition carries learning into new situations.
Research blog
Arm-wise Compositional Generalization in Dual-Arm Vision-Language-Action Models
A video-led introduction to ACG-Bench and AE-VLA: reusing arm-level skills under new orders, timing, and task combinations, from simulation to real robots.
Read the research updateExploring social interaction as a source of reward. Early experiments on learning from other agents, transfer beyond the simulator, and the limits of social feedback.
Read the research updateHow machines evolve
Our projects differ in scale and embodiment, but share a common mechanism: intelligence must keep changing through its consequences rather than remain fixed after training.
Generate experience by acting with agents, societies, tools, and the world.
Turn environmental and social feedback into better decisions over time.
Create new learning signals, curricula, and capabilities from the system’s own experience.
Reuse and recombine learned skills to solve situations not seen during training.
Digital World
The digital world gives agents a scalable environment for communication, social interaction, and rapid experimentation. OASIS lets us observe how behavior and intelligence change across populations of up to one million agents.
Physical World
The physical world grounds learning in space, bodies, and consequence. SPAgent connects multimodal reasoning to spatial tools, while Future develops compositional policies, benchmarks, and hardware infrastructure for embodied evolution.
A unifying principle
Evolution is the loop between agent, experience, and world.