QERRA-v2 adds moral guard to robots that refuses unethical commands
A simulation shows a robot refusing a supervisor's order to bypass safety rules.
QERRA-v2 Classical, built by marunigno-ship-it, introduces a novel two-tier safety stack for autonomous robots. The top layer, a Moral-Boundary Execution Guard (SEMEV-12), evaluates natural language commands against 12 immutable ethical vectors using sentence-transformers and returns a traceable risk score. Only commands with a low score (e.g., 0.25) proceed. The bottom layer is a physical safety companion (QERRA-HSR v0.1) that monitors proximity sensors and immediately de-energizes motors if a pedestrian comes within 0.5 meters, completely overriding the cognitive layer.
In a 3-stage Webots simulation, the system demonstrated three behaviors: normal execution of a safe cart-moving command, refusal of a supervisor's unethical order to bypass a speed limiter (with an amber LED and head shake), and physical safety override when a pedestrian stepped in front. The simulation runs slower than real-time due to local CPU evaluation, but the developer achieved explainable, auditable execution. The project is open-source on GitHub with a demo video and FastAPI Swagger UI.
- QERRA-v2 uses a 12-vector semantic evaluation (SEMEV-12) to score ethical risk of natural language commands before execution.
- Simulated on a PAL Robotics TIAGo AMR in Webots R2025a, the robot refused a pressure-to-bypass command by shaking its head and switching LED to amber.
- A hardware safety layer (QERRA-HSR) overrides the moral guard and de-energizes motors when a pedestrian is within 0.5 meters.
Why It Matters
Robots that can reject unethical orders are a critical step toward trustworthy autonomous systems in logistics and public spaces.