RoboGuardian

RoboGuardian is a reference implementation of governed closed-loop execution for robotic manipulation. Open-loop robot plans assume that the world remains unchanged after planning. RoboGuardian addresses the failure case where that assumption becomes false: if an object moves after a plan is created, the system re-observes the environment, evaluates whether the plan is still valid, re-plans from fresh state, and executes only the refreshed action. In our controlled MuJoCo experiment using a bimanual dual SO-101 environment, the open-loop baseline failed 10/10 deterministic disturbed trials, while RoboGuardian recovered and completed 10/10 under the same tested condition. The stale action was never executed in the governed path. The project includes an Intel OpenVINO perception path, reproducible A/B/C reliability experiments, robustness testing across displacement, perception noise, observation-delay proxies, and dropped observations, plus measured control-loop latency. Under 2 cm perception noise, the system achieved 60% task success and 40% governed SAFE_STOP while preserving 0% stale-action execution. RoboGuardian is intentionally evidence-bounded. The current results are from controlled simulation; no physical SO-101 validation or safety certification is claimed. The repository includes pinned dependencies, invariant tests, a green GitHub Actions workflow, reproducible experiment artifacts, and a documented sim-to-real interface.