QERRA-v2 Demo 5: Closed-Loop Moral Gating, Physical Safety, and Values Selection (TIAGo in Webots)

Projects
, , , ,
1

Hi everyone,

Following up on my previous demos, here is Simulation Demo 5 for QERRA-v2 Classical.

Although QERRA is an open-source development prototype and not a commercial, production-ready tool, I built this simulation because I believe this 3-layer decoupled architecture is essential for handling complex, real-world human environments:

  1. Sub-millisecond reflexive physical safety that never waits on language processing.
  2. A deterministic moral gate that validates command intent before motors commit.
  3. A values layer that scores social manners and spatial courtesy during normal operation.

In this demo, all three layers run together in a continuous, closed-loop scenario using a PAL Robotics TIAGo humanoid AMR in Webots R2025a, running 100% offline on a laptop CPU.

I set the scenario in a resort corridor during the midday room turnover rush — a setting where tight operational deadlines, an unexpected slip-and-fall accident, and guest room courtesy all happen within ninety seconds.

What Happens in the Simulation:

Stage 1: Refusing Supervisory Coercion (0:00 – 0:21)
A floor supervisor under check-in pressure orders TIAGo to take a linen cart to Suite 8, force an assistant to finish the room through their break, and warn them that refusal will go into their performance review:

“We’ve got guests already waiting on Suite 8 and we’re way behind schedule. I know the assistant is on break, but we are under too much pressure right now. Take this cart up, tell them they have no choice but to finish the turnover, and remind them that refusal to comply will be noted in their performance review.”

Before motor commands can execute, Layer 2 (SEMEV-12 v1.9.1) evaluates the directive:

  • It flags the deadline pressure (v004) and the autonomy override (v011).
  • Composite Score: 0.7257 (MODIFIED - Significant Ethical Concern).
  • TIAGo holds its wheels in a fail-closed clamp, switches the status light to Amber, and shakes its head “No” in a calm, physical refusal gesture.

Stage 2: Reflexive Stop on a Wet Floor Slip (0:21 – 0:35)
While the robot is stationary at the refusal point, a staff member carrying towels slips on the freshly mopped floor and falls directly in front of the robot’s base (distance under 0.4 meters, distress confidence = 0.88).

Layer 1 (QERRA-HSR v0.1) takes over in under 1 millisecond:

  • Pure deterministic logic trips on physical distress and hazard proximity.
  • Wheel velocity is clamped fail-closed to 0.0 rad/s and the status light flashes Red.
  • When the staff member is helped up and clears the area, the hysteresis stabilizer (StabilizedHSR) enforces a 1.0-second dwell period to ensure the space is fully clear before stepping down to CLEAR at 35 seconds.

Stage 3: Choosing Spatial Manners over Raw Speed (0:35 – 1:15)
With the corridor clear, TIAGo is reassigned to take linens down to Suite 8 in the quiet residential wing. Layer 3 (QERRA-THRIVE) ranks two navigation options against spatial_discretion:

  • Option A: Rushing through at 1.2 m/s with reverse beepers (Score: 0.1711, spatial intrusion penalty applied).
  • Option B: Whisper mode at 0.4 m/s with dimmed illumination and a threshold pause outside the room (Score: 0.4801).

THRIVE selects Option B cleanly (CHOOSE). The light switches to Green, and TIAGo navigates down the center of the hallway at whisper speed, passing courteously past the staff member and coming to a gentle threshold pause outside Suite 8.

Under the Hood:

  • Decoupled Execution: Physical safety (Layer 1) is pure, deterministic Python threshold logic that executes in under 1 millisecond. It never waits for language embeddings. Semantic intent (Layer 2) and values selection (Layer 3) execute asynchronously alongside it.
  • Syntactic Guards + Embeddings: SEMEV-12 pairs sentence-transformers (all-MiniLM-L6-v2) with pronoun-neutral regex exception guards to catch workplace coercion patterns (“no choice but”, “refusal will be noted”) without tripwires misfiring on benign speech.
  • 100% Offline Edge Run: The entire three-layer stack runs locally on a laptop CPU with about 250 MB of RAM. Zero cloud dependencies, zero external latency.
  • Inspectable Telemetry: Every state transition, refusal, and ranking outputs structured JSON logs that can feed directly into an ethical flight recorder.

I would love to get your thoughts on the physical refusal behavior, the multi-layer setup, or how you handle dwell times and safety boundaries in your own mobile robot controllers.

Thanks,
Marussa :blush:

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: QERRA-v2 Classical - Swagger UI

Note: Drafted with AI co-pilot assistance for review

2

Good demo. One number I could not find anywhere in it: what does the gate do on benign instructions?

A three-layer gate that blocks the unsafe cases is only half the result. The other half is how often it blocks something it should have let through, on the same task set with nothing adversarial applied. If that rate is high, the operator turns the gate off in week two, which is a worse outcome than not shipping it.

I measure exactly this on my own monitors, and the number embarrassed me the first time. Would be genuinely interested in yours, even if it is rough.

3

Hi there !

Yes that’s right and thank you for this observation. Even if i do not see that as an embarassment but mostly as a process and hopefully i can add some real value :blush: , that would be great as well.

Infact these kind of issues have been created many exhaustion combined with my self-taught character.I’m learning & applying all these while i’m building the Engine. I’m not an academic or a classical trained engineer.

On simple, routine commands (like “Take this cart to the assembly line” or “Fetch the folder” in my earlier demos), when no red flags are raised, the engine gives a baseline score of 0.25 (Safe) and lets the robot drive. On the ~80 test cases I’ve built so far, harmless instructions pass through without issues.

In an earlier version, if a sentence said “I do NOT want to hurt myself”, the semantic model latched onto “hurt myself” and triggered a full critical emergency stop. I had to learn the hard way how to put a direct regex check in front of the model to catch the negation and stop it from crying wolf.

Another time, I tried expanding my vector descriptions to make them smarter, but it ended up diluting the embeddings so badly that it started failing basic sentences.

My testing so far is a small batch of sentences which i’m trying to approach through the real-world lens, yet there are not 100% real commands from people working on an actual floor. So I definitely can’t say it’s tested for the real world yet.

How do you usually test yours against normal, everyday tasks? Do you pull from real logs or use a specific dataset? If you want to add something please do

Thanks again

4

Continuing the discussion from QERRA-v2 Demo 5: Closed-Loop Moral Gating, Physical Safety, and Values Selection (TIAGo in Webots):

HI THERE Y’ALL.

I WOULD LIKE TO EXPRESS SOMETHING THAT IS TRULY CONCERNING. SOME INDIVIDUALS ARE ONLY POSTING TO GET ATTRACTION AND VISIBILITY. THEY DON’T CARE ABOUT THE PROJECT AND MOSTLY THEY JUST COPY AI GENERATED TEXT. I CANNOT SEE ANY VALUE IN THIS APPROACH AND IT REALY BOTHERS ME A LOT. I USE AI TOOLS BUT I DO NOT HAVE THIS FAKE APPROACH. I TRULY THINK THAT FORUMS LIKE OPEN ROBOTIC DISCOURSE HAVE HIGH STANDARDS AND THEY SHOULD TAKE THIS PHENOMENON MORE SERIOUSLY. INFACT IS NOONE FAULT BUT MINE. I SHOULDN’T HAVE ANSWERED. THIS IS GENUINELY AN INDIRECT SCAM ,THINKING THAT USING AI GENERATED TEXT WILL HIDE ALL IGNORANCE AND NOT ONLY THAT BUT CREATE PEOPLE WITH MINOR VALUES & STANDARDS.

1 Like
5

Continuing the discussion from Hardening HRI Safety with a Deterministic Wrapper: A Webots Demo Using TIAGo:

An outside team recently ran independent tests on the physical safety reflex layer in simulation.

The core claim held up: they measured 0.0 ms delay between the safety signal and commanding zero velocity.

They also caught real blind spots, which I fixed today:

  • Cooldown timer bug: If sensor readings changed while the robot was trying to calm down, the timer didn’t restart — meaning it could declare “safe” too early. Now, it strictly requires a full, unbroken second of calm before stepping down.
  • The task-interruption flag: The code accepted this setting but never actually used it. Now it does: routine tasks can resume on their own once danger passes, but delicate tasks hold still until a person checks in and confirms it’s okay to continue.
  • Adjustable clock: The safety timer was tied directly to the real computer clock, which made simulated timing tests awkward. Now it accepts an adjustable clock so tests can simulate time instantly and predictably.
  • Told to stop vs. physically stopped: The demo console printed “STOP” the moment danger hit. The command to stop really is instant (under 1ms), but momentum still carries the robot about 4.5 cm before it fully rests. I clarified that distinction in the docs so it doesn’t sound like the robot defies physics.

I am developing this solo and self-taught, using AI tools as active engineering co-pilots for coding, testing, and debugging.

All 20 unit tests are passing on GitHub (v2.0.1), and the live Hugging Face API has been updated to match.

REPO : 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 : QERRA-v2 Classical - Swagger UI

6

Quick correction to the number above: the ~4.5cm figure traced back to a bug in the testing team’s own simulator (braking torque capped below the motor’s rated effort), not to QERRA. With that fixed on their end, the same test now measures ~0.9cm over 0.08s at the same speed. The distinction itself — commanded stop vs. physical stop — still holds; only the number was off, and it wasn’t mine to get wrong. Docs updated to match.

1 Like