I study how agents form internal worlds before they act in the physical one.

causal representation learning, intervention-driven reasoning, physical understanding

• Why it matters: better robustness, sharper explanations, and fewer brittle shortcuts.
• Representative signals: Causal-PIK and recent causal world-model work aimed at physical reasoning.

decision-grounded world models, latent imagination, counterfactual planning

• Why it matters: an agent should imagine alternatives before paying the price in the real world.
• Representative signals: Dreamer-style imagination, JEPA-style abstraction, and evaluation work such as WorldGym.

3D reasoning, scene representations, spatial cognition

• Why it matters: a capable robot should reason over space, not only over pixels or language tokens.
• Representative signals: recent work on 4D spatial intelligence and multi-image spatial reasoning benchmarks.

model-based RL, safe exploration, sim-to-real reliability

• Why it matters: real systems are expensive, risky, and too slow for wasteful training loops.
• Representative directions: learned environments for control, action-constrained RL, and policy evaluation before deployment.