Teaching AI to understand how objects change and respond

Current world models either generate videos or reconstruct scenes, but don't explicitly represent how objects behave and change state. WorldString learns a state manifold for real-world objects directly from point cloud or RGB-D video input, treating objects as actionable entities with intrinsic properties that determine their behavior. The fully differentiable architecture integrates seamlessly with policy learning and neural dynamics models, positioning it as a foundational primitive for building physical world models that understand object interactions.