Physical Reasoning

The field asks whether AI can infer how objects and systems behave from images, videos, diagrams, and sensorimotor interaction, then use that understanding to predict outcomes or choose actions. Approaches include physics-informed models and simulators, structured world models that represent physical states, and vision-language model training with stepwise rewards, demonstrations, or interactive exploration. Evaluations range from physics problems and synthetic collision or mechanism tasks to video prediction, robotic manipulation, and interactive environments, with increasing attention to counterfactuals and out-of-distribution transfer. A central open problem is distinguishing grounded causal understanding from reliance on language or visual shortcuts: models still struggle with visual evidence, complex dynamics, and reliable long-horizon action.