Everything about Physical AI Governance

1. Key terms

Governance terms used throughout this guide. Technical terms are in the Hardware and Software guides.

2. The four-layer model

Governance is the foundation layer. Everything above it (hardware, software, complete systems) operates within the boundaries that governance defines. Without governance, a robot arm is a liability lawsuit waiting to happen, an autonomous car is a weapon, and a surgical robot is a medical malpractice case.

3. Why governance is the least developed layer

Hardware is mature (servo motors have not fundamentally changed in decades). Software is advancing rapidly but has clear frameworks (ROS2, Isaac, LeRobot). Governance is where Physical AI has the biggest gap between what exists and what is needed.

What governance exists today

Robot safety standards (ISO 10218) for industrial arms. Vehicle safety rules (NHTSA / FMVSS) for cars. Medical device approval (FDA 510(k)) for surgical robots. EU AI Act for high-risk AI. These are sector-specific, well-established, and functional. They work for the systems they were designed for.

What governance is missing

No framework governs autonomous AI agents acting in the physical world. No standard for auditing world models (Cosmos, JEPA). No protocol for inspecting AI intent before physical action. No liability framework for when an AI agent autonomously causes harm. No governance for humanoid robots operating in human spaces. Only 14.4% of organizations have security controls for AI agents.

4. The regulatory landscape

Physical AI governance comes from multiple sources, each covering different aspects. No single authority governs the whole stack.

5. EU AI Act and Machinery Regulation

The EU has the most comprehensive governance framework for Physical AI. Three major regulations apply simultaneously to AI-powered robots.

EU AI Act (Regulation 2024/1689)

Entered into force: August 1, 2024. Phased enforcement.

Core concept: risk-based classification. Not all AI is regulated equally. Systems posing higher risk face stricter requirements.

For Physical AI specifically: an industrial robot with AI-powered safety functions (human detection, collision avoidance) is automatically classified as high-risk because it falls under both the AI Act AND the Machinery Regulation.

AI Act Omnibus (agreed May 7, 2026)

The EU Council and Parliament reached political agreement on significant amendments to the AI Act under the Omnibus VII simplification package on May 7, 2026.

Formal adoption expected before August 2, 2026. Sources: EU Council press release; AI Act tracker.

EU Machinery Regulation (2023/1230)

Published June 2023, replacing the old Machinery Directive 2006/42/EC. The key change: it explicitly addresses AI in machinery. If a robot uses machine learning for a safety function (detecting humans, avoiding collisions, adaptive control), it triggers high-risk classification and requires third-party conformity assessment.

Why this matters: under the old Directive, most machinery could self-certify. The new Regulation says: if your robot's safety depends on AI, an independent body must verify it. A fundamental shift from "we checked ourselves" to "someone external checked us."

For Physical AI: nearly every robot with computer vision or adaptive behavior falls under this. A cobot that uses YOLO to detect human presence is using AI for a safety function. That triggers third-party assessment.

Revised Product Liability Directive (December 2026)

Adopted 2024, applicable from December 2026. The most significant change: software is formally recognized as a product for liability purposes. Under the old 1985 directive, only physical products were covered.

Now: if a piece of software is essential for the functioning of a robot, or can cause harm independently, it is subject to the same liability rules as physical products. This closes the "software is not a product" loophole that existed for decades.

Implications for GR00T, Cosmos, JEPA: when an autonomous robot uses a world model to predict consequences and acts on that prediction, and the action causes harm, the world model software is now a liable product. The manufacturer, the deployer, and potentially the model provider could all face liability claims.

6. NIST frameworks and the AI Agent Standards Initiative

NIST provides voluntary frameworks that heavily influence US industry standards and federal procurement requirements. Unlike the EU's regulatory approach, the US relies on voluntary compliance with NIST frameworks, supplemented by sector-specific regulations (NHTSA for vehicles, FDA for medical).

AI Risk Management Framework (AI RMF 1.0)

Released January 2023. Provides a structured approach to identifying, assessing, and mitigating AI risks. Organized around four functions: Govern (establish policies and accountability), Map (identify and characterize risks), Measure (quantify risks with metrics), Manage (prioritize and act on risks).

For Physical AI: NIST released a concept note on April 7, 2026 for an AI RMF Profile on Trustworthy AI in Critical Infrastructure. This guides operators of critical infrastructure (power plants, transportation, healthcare) on specific risk-management practices when deploying AI.

Key insight: the AI RMF is not a compliance checklist. It is a thinking framework. It forces organizations to answer: "What could go wrong? How do we know? What do we do about it?" This contrasts with the EU's prescriptive approach (specific requirements listed in the AI Act).

NIST AI RMF

AI Agent Standards Initiative (February 2026)

Launched February 17, 2026 by NIST's Center for AI Standards and Innovation (CAISI). The first US government program dedicated to developing standards for autonomous AI agents. Covers agents that can act autonomously, use tools, interact with APIs, and execute multi-step tasks.

NCCoE concept paper: separately, NIST's National Cybersecurity Center of Excellence published a concept paper on AI agent security. Key insight: it treats prompt injection not as a model-quality issue but as a security control problem. Prevention and mitigation need to be designed into the architecture, not patched on after deployment. Comments closed April 2, 2026.

SP 800-53 control overlays: NIST is developing security control overlays for both single-agent and multi-agent AI systems, building on the existing SP 800-53 framework.

Critical gaps (as of May 2026): no standalone federal agentic AI security standard exists. No FAR (Federal Acquisition Regulation) clause governs AI agent procurement. MITRE ATT&CK does not cover agentic attack patterns (multi-agent lateral movement, reasoning-layer manipulation). Only 14.4% of organizations have security controls for AI agents.

NIST announcement

NIST Measurement Science for Robotics

NIST's Intelligent Systems Division runs the Measurement Science for Robotics and Autonomous Systems Program. It develops test methods and performance metrics for: safe operations of collaborative robots, effectiveness of dexterous manipulators, mobility and safety of mobile robots, perception accuracy in dynamic environments, human-robot and robot-robot interaction safety, robot agility in unpredictable environments, and validation of AI/ML using well-documented datasets.

This is the practical counterpart to the policy frameworks: how do you actually test whether a robot is safe? What does "safe enough" mean in measurable terms?

7. Robot safety standards (ISO)

ISO standards are the international backbone of robot safety. They define what "safe" means in measurable, testable terms. Regulations are laws enforced by governments; standards are technical specifications that regulations reference. The EU Machinery Regulation says robots must be safe; ISO 10218 says what "safe" means in practice.

8. Autonomous vehicle regulation

The most active governance battleground. Autonomous vehicles operate on public roads where errors kill people. The US, EU, and China are racing to establish frameworks, but none has a comprehensive federal law yet.

US: NHTSA framework and FMVSS modernization

Current state: NHTSA provides voluntary guidance for SAE Levels 3-5. Federal Motor Vehicle Safety Standards (FMVSS) are being modernized to allow vehicles without human controls.

FMVSS rulemaking (2026): NHTSA proposed amendments to FMVSS No. 102 (transmission shift position), FMVSS No. 103 (windshield defrosting), and FMVSS No. 104 (windshield wiping) on March 16, 2026. The amendments except vehicles without manually operated driving controls from requirements designed around a human driver. Four more proposed rulemakings are planned in 2026 covering indicators, telltales, warning lights, and lighting in ADS-only vehicles.

AV STEP program: the AV STEP voluntary oversight program (proposed January 2025) creates dedicated exemptions for AV deployment.

SELF DRIVE Act of 2026: proposed in Congress (third attempt after 2017 and 2021). Would be the first federal statute dedicated to AV safety. Key provisions: requires manufacturers to develop a safety case (structured argument that the system is safe), creates a federal crash-reporting repository, and prevents a patchwork of state regulations.

The liability question: as vehicles move from Level 2 (human supervises, Tesla FSD today) to Level 4 (no human needed, Waymo in geofenced areas), liability shifts from the human driver to the ADS provider. When a Level 4 vehicle crashes with no human in the loop, who is liable? The vehicle manufacturer? The ADS software provider? The owner who chose to use it?

EU and international

UNECE UN Regulation 157: the first international type-approval standard for Level 3 systems (Automated Lane Keeping Systems). Limits speed to 60 km/h. Requires driver monitoring. Adopted in Japan, the EU, and other UNECE members.

UNECE Global Technical Regulation (January 2026): new harmonized methodology for validating ADS-equipped vehicles. Safety case approach. Anchored in robust R&D processes. Represents global convergence on safety validation methodology.

China (April 2026): Ministry of Industry and Information Technology proposed mandatory safety standards for intelligent connected vehicles. 62-page proposal. Public comments closed April 13, 2026. China aims for nationwide AV regulation and mass production of humanoid robots by 2027.

9. Medical and surgical regulation (FDA)

Medical devices face the strictest governance of any Physical AI category. The FDA's regulatory pathways ensure that surgical robots, exoskeletons, and AI-assisted diagnostics are safe before they touch patients. This is the sector where governance is most mature.

10. Industry governance: NemoClaw and safety-by-design

Industry is not waiting for regulators. Companies are building governance into their products because customers (enterprise buyers) demand it, and because liability risks are real.

NVIDIA NemoClaw (GTC 2026)

NemoClaw is NVIDIA's security and governance layer for AI agents, announced at GTC 2026. It works alongside AgenticROS and the broader NVIDIA Isaac platform. NemoClaw monitors the AI agent's reasoning process in real time and enforces safety guardrails.

The key phrase: NemoClaw "inspects the intent of the AI's logic." This is not monitoring actions after they happen (reactive). It is monitoring reasoning before actions execute (proactive). This is a fundamental distinction and the first commercial product that attempts to govern AI intent, not just AI output.

What it monitors: AI agent reasoning chains, context understanding, tool use decisions, and cross-system interactions. If the agent's reasoning violates safety constraints, NemoClaw intervenes before the physical action occurs.

Limitation: NemoClaw is a security tool, not a governance framework. It enforces technical guardrails defined by the deployer. It does not answer the higher-level questions: who decides what the guardrails should be? Who audits the guardrail definitions? Who is accountable when guardrails fail?

Safety-by-design practices

11. Six governance gaps

These are the areas where governance does not yet exist but is urgently needed. The biggest risks and the biggest opportunities for governance practitioners.

Detailed gap analysis

• Autonomous AI agent liability. When an AI agent acts autonomously and causes harm, who is liable? The user who delegated authority? The company that deployed the agent? The vendor that built the model? No jurisdiction has answered this clearly. Amazon v. Perplexity (November 2025) is an early test case.
• World model governance. Neither Cosmos (NVIDIA) nor JEPA (LeCun) has any governance framework. A world model predicts physical consequences and informs autonomous action. Who audits these predictions? How do you verify a latent-space representation of physics?
• Humanoid robot governance. No governance framework exists specifically for humanoid robots operating in human spaces (homes, hospitals, airports). ISO 13482 covers "personal care robots" but was written before modern humanoids existed. JAL deployed humanoids at airports (May 2026) with no humanoid-specific safety standard.
• Cross-border AI agent interoperability. NIST mentions MCP as a candidate protocol, but no international agreement exists on how AI agents should communicate, authenticate, or be governed across jurisdictions. EU and US have fundamentally different approaches (prescriptive vs voluntary). A robot operating in both needs to satisfy both.
• AI intent inspection. NemoClaw "inspects intent" but is a proprietary, single-vendor tool. No open standard exists for inspecting AI reasoning before physical action. No interoperable protocol for intent auditing across different AI platforms.
• Agentic attack surface. MITRE ATT&CK does not cover agentic attack patterns. Multi-agent lateral movement, reasoning-layer manipulation, cross-system context poisoning. The security community's standard threat model has not caught up to agents that can act autonomously.

12. The Intent Layer thesis

The Intent Layer thesis argues that understanding, not tools, is the core differentiator in human-AI collaboration. Applied to Physical AI governance:

NemoClaw is the first commercial product that moves toward intent inspection. The NCCoE concept paper treats prompt injection as an architectural problem, not a model quality problem. The NIST AI Agent Standards Initiative asks for audit trails of agent reasoning. These are all early signals of the same direction: governance is moving from monitoring outputs to inspecting process.

Physical AI makes the argument concrete: when a humanoid robot reaches for an object, the governance question is not just "did it exceed 150N of force?" but "why did it reach for that object in the first place, and was that intent aligned with what the human actually wanted?"

13. Regulatory timeline

What is happening when. This is the fastest-changing layer. Verify dates before acting on them.

14. Resource directory

EU regulation

• EU AI Act tracker (full text)
• EU Machinery Regulation (EUR-Lex)
• EU Council press release on the Omnibus (May 7, 2026)

US frameworks

• NIST AI RMF
• NIST AI Agent Standards Initiative
• NIST Robotics Program
• NHTSA Automated Driving Systems

ISO and OSHA

• A3 Association for Advancing Automation (ISO 10218 distribution)
• OSHA robotics standards

Industry governance

• AgenticROS (the bridge that NemoClaw works alongside)
• ROSClaw paper (arXiv)