Intervention-Based Movability Representations in Minimalist Embodied Predictive Agents

This preprint studies whether minimalist embodied predictive agents can form movability-sensitive internal representations from intervention-driven experience. The work uses a 2D embodied world, vector causal perturbation diagnostics, causal-swap tests, blind generalization, temporal and noise perturbations, visual shortcut tests, distillation, and 3D stress tests. The central result is cautious: shortcut-deprived sensor agents form stable movability-like representations, while raw visual and 3D settings expose major unresolved limits.