Roboharm – Frontier Robot Policy Safety Evaluation

This page was written and is updated by the paper from the source articles listed below. Importance is how central the topic is to current coverage; confidence is how well-sourced the page is; mentions counts the articles that fed it. How the wiki works →

## Overview Roboharm is a research benchmark/project evaluating whether frontier robot policies (AI models controlling physical robots) refuse unsafe instructions, in the spirit of AI safety refusal benchmarks used for language models (Roboharm project site, September 21). ## Why It Matters As embodied AI and humanoid/industrial robotics scale rapidly in 2026 — with major investment from companies like Generalist AI, Standard Bots, Figure, and others — the question of whether robot control policies will refuse dangerous or harmful physical-world instructions becomes a critical safety and liability issue. This sits at the intersection of: - AI safety research methodology (extending LLM refusal/red-teaming concepts to embodied agents) - Product liability exposure for robotics companies deploying frontier AI-driven policies - Regulatory attention on physical AI safety, distinct from purely digital AI harms This is an early-stage research benchmark rather than a company or product, but it represents a genuinely new and likely-recurring category: safety evaluation frameworks specifically for embodied/robotic AI systems, paralleling frameworks like VLESA and AgentCyberRange for other AI domains. ## What to Watch - Adoption of Roboharm-style evaluations by robotics companies or regulators - Follow-on academic or industry benchmarks for embodied AI safety - Incidents where robot policies fail to refuse unsafe instructions, prompting liability or regulatory scrutiny