TourPhysics: Bringing Physics to World Models for Exploration and Manipulation from a Single Image
2026-09-07 12:00Models🔥 42.2 heat score
1sources
1days unfolding
42.2heat score
2mentions
SummaryAI generated
TourPhysics is an online framework based on single images and declarative physical configurations, aimed at enabling persistent exploration and manipulation tasks by combining deterministic simulation with video generation technologies. This framework assigns independent roles to simulator state, geometric evidence, generator control, and appearance memory: for each action, the simulator first calculates finite physical and camera trajectories, followed by the generation of corresponding observations; the accepted observations update the terminal state and appearance memory, while the committed states and simulator geometry remain fixed during synthesis and retrying. Additionally, TourPhysics separates simulator geometry used for projection visibility from relative depth used for conditional generation, and retrieves static appearances across views using reference anchored memories when no effective match is found, returning to the original path in such cases. Experiments show that, compared to baseline methods, TourPhysics approaches preset trajectories more closely, maintains the integrity of the input scene, and reduces appearance drift during long-range revisit periods.