Can Predicted Dynamics Exist in the Physical World?

Researchers propose a physical-admissibility gate for AI-generated robot plans that filters out invalid actions before execution.

The gate checks state rollouts, action chunks, and latent plans using kinematic, dynamic, and horizon-based conditions.

On the Hugging Face LeRobot PushT benchmark, it rejected most invalid proposals while preserving near-optimal progress.

The method outperformed kinematic-only checks, showing that plans can appear good but still be physically impossible.