Reliable rover autonomy requires a mechanism for rejecting unsafe command sequences before they reach the vehicle. This paper presents a command-level validation architecture built around a live, scene-aware digital twin. A rover-mounted RGB-D sensor supplies depth and color data to NVIDIA nvblox, which reconstructs the local environment as collidable geometry. NVIDIA cuVSLAM estimates the rover pose, and the synchronized rover and scene are loaded into NVIDIA Isaac Sim. Candidate wheel-command sequences are replayed in physics and evaluated for feasibility, collision, obstacle clearance, and endpoint error before ROS 2 dispatch. Route generation remains outside the safety boundary; an LLM-based generator was used only to produce nontrivial candidates for evaluation. A controlled 100-trial simulation battery showed complete agreement between the gate and analytical labels for clear and blocked direct-path controls. Among generated candidates, physics playback identified 34 routes that passed the waypoint-level geometric screen but subsequently stalled or became wedged during simulated execution. Two matched hardware campaigns then evaluated agreement between simulated and tape-measured rover motion. Across eight tested motion conditions, average simulated-to-measured endpoint-position error decreased from 2.9 cm before calibration to 0.5 cm after motor recalibration. These results support live physics rehearsal as both a practical command-safety layer and a diagnostic tool for improving rover calibration.

Keywords

Digital Twin, Simulation-Based Validation, Lunar Rover, Scene Reconstruction, and Mobile Robotics.