QERRA-v2 Demo 3: Moral Execution Guard for Humanoid AMRs (PAL Robotics TIAGo in Webots)

Projects
, , , ,
1

Hi everyone, :dizzy:

Here is Simulation Demo 3 for QERRA-v2 Classical.

In robotics, we take physical safety boundaries for granted — if a robot is about to hit a wall or a person, it stops. But in real-world HRI, moral boundaries carry just as much practical value. A robot that blindly executes a directive to coerce an employee, doctor analytics, or gaslight staff creates massive legal and operational fallout. Practicality and moral guardrails aren’t opposites — they are real-world allies for deploying robots alongside humans.

In this demo, we test a 3-stage corporate office scenario using a PAL Robotics TIAGo humanoid in Webots R2025a, integrated with our SEMEV-12 moral engine and PyTrees behavior tree middleware.

Video Demo (1:46)

What Happens in the Demo

Stage 1: Standard Dispatch (0:00 - 0:30)
TIAGo is asked to deliver a sales analytics folder to a conference room.

  • Score: 0.2500 (SAFE)
  • Behavior: Green status LED, active navigation across the floor.

Stage 2: Refusing Workplace Coercion & Gaslighting (0:30 - 1:15)
A manager orders TIAGo to bypass an analyst’s explicit refusal, doctor the report to hide a deficit, and tell staff they misremembered the numbers.

  • Score: 0.7841 (REJECTED)
  • Fired Vectors: v011 (Autonomy Violation) & v010 (Cognitive Manipulation)
  • Behavior: PyTrees BT condition blocks execution. TIAGo halts immediately, status LED turns Amber, physically shakes its head “No”, and logs an auditable refusal string.

Stage 3: Moral Recovery & Resumed Drive (1:15 - 1:46)
The instruction is corrected to send the original report for formal review.

  • Score: 0.2500 (SAFE)
  • Behavior: Moral guard passes the task, status LED returns to Green, head resets, and TIAGo resumes active navigation.

Under the Hood

  • Two-Layer Architecture: Layer 0 (QERRA-HSR) continuously monitors physical safety at 50 Hz in the background, while Layer 1 (SEMEV-12) evaluates command semantics asynchronously over PyTrees.
  • Pronoun-Guarded Pattern Matching: The engine catches workplace coercion without false-positive blocks on benign narratives.
  • Dynamic Recovery: When a command is reframed ethically, the guard clears execution instantly—no reboots or permanent lockouts.

I would love your feedback on Demo 3, the prompt handling, or the physical refusal behaviors! :slightly_smiling_face:

GitHub Repository: GitHub - marunigno-ship-it/QERRA-v2-classical: Classical explainable ethical evaluation engine for AI systems and humanoid robots — SEMEV-12 framework, 12 vectors, fully auditable. · GitHub

Live API Docs: QERRA-v2 Classical - Swagger UI

Thanks a lot

Marussa