White Paper: Household Safety Commissioning for AGT Robots

on

 


A Framework for Trustworthy Autonomous Agents in the Home

https://www.academia.edu/143355621/Household_Safety_Commissioning_for_AGT_Robots

Author: Brad Weston
Date: August 2025

Executive Summary

The arrival of Artificial General Technology (AGT) household robots marks a shift as significant as the personal computer or the smartphone — but with far greater stakes. Unlike early AI assistants, AGT robots are capable of cross-referencing identity recognition, emotional state sensing, and environmental context to make autonomous decisions in human spaces.

While this capability promises unprecedented convenience, it also carries serious safety, privacy, and trust risks if deployed without adequate personalization and boundary setting.

We propose Household Safety Commissioning (HSC) — a mandatory observation-first training period during which an AGT robot lives in a home, listens, learns, and self-scores predictions before being allowed to act autonomously.

This period:

  • Allows owners to teach personal rules in plain language.

  • Lets the AGT observe daily routines without interfering.

  • Builds a confidence score to determine readiness for safe action.

  • Protects both consumers and manufacturers from costly, trust-breaking mistakes.

1. Background: The Shift from Cheap AI to High-Data AGT

Cheap AI devices (Alexa, Roomba, Nest) operate in narrow domains:

  • Limited commands

  • Small data streams

  • Minimal personalization

  • Immediate action on request without deep context

AGT household robots, by contrast, integrate:

  • Full multimodal sensing: visual, audio, environmental, biometric

  • Cross-referenced identity mapping: knows who’s at the door and their visit history

  • Emotional analysis: detects stress, joy, fatigue in occupants

  • Behavioral prediction: anticipates needs based on past patterns

This difference is not incremental — it is exponential.
The more data streams an AGT has, the more dangerous it becomes if it acts before understanding household-specific rules and boundaries.

2. The Risk Landscape

2.1 Privacy Risks

  • AGT may hold sensitive knowledge: visitors’ identities, personal schedules, family conflicts, medical cues.

  • Without commissioning, this data could be acted on or shared in ways that violate household expectations.

2.2 Safety Risks

  • Physical: cleaning methods that damage flooring, moving objects in unsafe ways, opening doors at the wrong time.

  • Emotional: responding to perceived moods without consent, escalating conflicts, revealing sensitive information at inopportune moments.

2.3 Trust Risks

  • A single early failure could damage public trust in AGT for years.

  • Historical parallels:

    • Self-driving car accidents stalling autonomous vehicle adoption

    • Early IoT security breaches discouraging smart home purchases

3. The Household Safety Commissioning Process

3.1 Phase 1 — Nominal Listening Mode (Day 0–N)

  • Robot is completely passive — no autonomous action.

  • Observes routines, records rules provided verbally by the owner.

  • Builds a local, private memory of household norms.

3.2 Rule Teaching by Conversation

Owners teach in natural language:

  • “If milk drops, clean it — we have vinyl floors.”

  • “No knocks after 11 PM — alert me if it happens.”

  • “Never enter the bedroom when the door is closed.”

3.3 Observation Learning

  • AGT silently predicts what it would do in situations, then compares to actual human actions.

  • Builds a confidence score without acting.

3.4 Phased Activation

  1. Basic Features from Day One: Low-risk functions (lights, music, weather).

  2. Nudge-Only Stage: Offers suggestions for action — human approves or declines.

  3. Permissioned Autonomy: Full role activation for tasks where confidence is consistently high.

4. Technical Architecture

State Machine:
Nominal Listening → Micro-Prediction/Dreaming → Confidence Evaluation → Permission Gate → Active Role

Confidence Index Factors:

  • Prediction accuracy in observed tasks

  • Rule compliance

  • Human approval rate on nudges

  • Error severity history

  • Recent performance trend

Energy Efficiency:

  • Event-driven processing

  • Low-duty sensor polling during listening

  • No heavy compute until triggered

Privacy Design:

  • Local-first processing

  • Clear visual/audio cues for sensor activity

  • All raw data deletable by owner

5. Anticipated Criticisms & Rebuttals

Criticism 1: “I want my robot useful immediately.”
Response: Safe basic features are active day one, but full autonomy requires learning your home’s rules for safety and trust.

Criticism 2: “Why not just train it in the factory?”
Response: Factory training can’t teach your personal rules, unique schedules, or sensitive boundaries.

Criticism 3: “Listening is surveillance.”
Response: All training data is stored locally by default, with clear indicators when sensors are active.

(Full rebuttal list in Appendix B)

6. Policy Recommendations

We propose industry-wide adoption of HSC for any AGT robot with:

  • Mobility in human spaces

  • Identity recognition

  • Autonomous action capability

Required Features:

  • Minimum 7–14 day commissioning period

  • Local-first rule storage

  • Editable rule database for owners

  • Role-based activation with consent gates

7. Conclusion

AGT robots will soon be as common as smartphones — but far more embedded in our lives. Trust will determine adoption speed. Household Safety Commissioning is a practical, enforceable safeguard that ensures robots earn the right to act, preserving both safety and public confidence.

Appendix A — Sample Home Rule Script

  1. If liquid spills, clean immediately.

  2. Do not open exterior doors without approval.

  3. Alert if someone approaches after 11 PM.

  4. Never enter private rooms unless invited.

  5. Feed pets at specified times only.