kinematic-conditions

This paper proposes a method for validating physical admissibility of AI predictions by using a prediction-control interface with kinematic and dynamic conditions. It demonstrates high accuracy in filtering invalid proposals on the LeRobot PushT task.