Ethyka robotics.md: A Verifiable Ethics-as-Code Specification for Physical AI Agents
Year: 2026
Type: preprint
Abstract: Roboethics has produced mature bodies of principles — human well-being, responsibility, transparency — but these remain high-level statements that are difficult to verify on a deployed robot. Meanwhile, foundation-model-enabled robots introduce a new failure mode: an agent whose plan can be manipulated through natural-language or network inputs into physically unsafe behavior. We present Ethyka robotics.md, an open "ethics-as-code" specification that addresses this gap for physical AI agents. The specification expresses eleven rules (ROB-01..11) — safety-first stopping, actuation limits, stop and override, movement honesty, sensor privacy, resistance to external manipulation, safe separation, operational envelope, forensic logging, safe exploration, and continuous hazard analysis — as machine-readable clauses with stable identifiers, measurable parameters, severities, and one adversarial test per rule. Numerical limits are not asserted by the specification; they remain parameters bound to the governing standards (ISO 10218, ISO/TS 15066, IEC 61508, ISO 13855, ISO 21448). A generation pipeline validates declared values against a JSON Schema and emits runtime guardrails, system prompts, a test suite, and a traceability manifest, with severity-based deployment gating. We describe the design, its grounding in the safety literature, and its limitations as a proposal that complements — not replaces — certification.
Author: Diezma, Pedro
Language: English
Subfield: Safety Research
Field: Social Sciences
Domain: Social Sciences
Open Access status: gold
OpenAlex ID: W7204545908
Cites:
Cited by:
Related to: