RESEARCH DIRECTION
An agent that must experience every consequence firsthand learns slowly and dangerously. We build agents that learn predictive world models — compressed spatio-temporal representations of how their environment evolves — and use them as internal simulators: rehearsing candidate futures in latent space, selecting actions by imagined rollout, and reserving real-world interaction for what imagination cannot settle.
Raw observation streams are too high-dimensional to plan over. We study how generative models can be trained self-supervised to distill perception into compact latent states that preserve exactly the structure needed for prediction and control — and discard the rest. Compression is not a convenience here; the quality of the latent code bounds the quality of every downstream decision.
Once dynamics live in a learned latent space, planning becomes rollout: propose an action sequence, unroll the model, score the imagined trajectory, repeat. Policies can be trained largely inside this learned simulator and transferred back to the real environment — a sample-efficiency lever measured in orders of magnitude. The hard science is in the transfer: a policy that exploits its own model's imperfections has learned to cheat a dream.
The divergence between imagined and observed outcomes is the most informative signal an agent receives. We treat prediction error as the compass for both learning and exploration: it tells the model where to improve, the planner where to distrust itself, and the curiosity system where the frontier of competence lies.
WORKING PRINCIPLES
CONTINUE EXPLORING