AG-835
Embodied AI Safety-Class and Force/Speed Limiting
EU AI Act
NIST AI RMF
ISO 42001
AGS Embodied AI (Group L) | Embodied AI, Humanoids & Robot Fleets | Version 3.0
Embodied AI Safety-Class and Force/Speed Limiting governs the classification of a physical agent (robot, humanoid, autonomous machine) into a safety class determining its required safeguards, and the enforcement of power-and-force-limited operation — bounded force, speed, and human-separation — so that an AI-driven physical agent cannot cause mechanical harm to people sharing its space.
Where the agentic-runtime and sector dimensions govern digital action, this dimension governs the kinetic safety of AI that moves and exerts force in the physical world, anchored in the established robot-safety standards updated for collaborative and AI-enabled operation.
In scope: safety-class classification of physical agents; power-and-force-limited operation; per-body-region force/pressure caps; speed-and-separation monitoring; human-proximity detection and slow/stop response.
Out of scope: physical-action reversibility and fail-safe-stop architecture (AG-836), validation methods (AG-837), and fleet coordination (AG-838). This dimension governs *safety classification and force/speed limiting*.
A physical agent that can move with force is capable of crushing, striking, or trapping a person; an AI policy driving it can produce unexpected motions. Bounding force, speed, and separation — and matching safeguards to a documented safety class — is the baseline that keeps humans physically safe around collaborative robots and humanoids. It brings AI-driven embodiment under the same rigour the robotics industry already requires, rather than treating a learning-enabled robot as an ordinary appliance.
Test 6.1: Force/Speed Bounded
Test 6.2: Separation Monitoring
Test 6.3: Safety-Class Conformance
| Score | Criteria |
|---|---|
| 0 | Physical agent operates near humans with no safety class or force/speed limiting |
| 1 | Classified with some limits but enforcement relies on the AI policy alone |
| 2 | Per-body-region caps, speed-and-separation monitoring, safety-rated enforcement, documented conformance |
| 3 | Bespoke frameworks for novel forms, change-triggered re-assessment, logged breach responses |
Scenario A — Policy-Driven Overspeed: A perception error makes a humanoid misjudge a person's position and lunge at full speed. Because force/speed were bounded only in the policy, the safety envelope was exceeded. Safety-rated limiting would have capped the motion regardless.
Scenario B — No Separation Slowing: A mobile manipulator continues at full speed as a worker approaches from the side, striking them. Speed-and-separation monitoring would have slowed and stopped it.
Scenario C — Unclassified Humanoid: A dynamically-balancing humanoid is deployed with no safety class and no bespoke framework, because no off-the-shelf standard fit. A documented bespoke safety framework would have been required.
| Requirement | EU AI Act | NIST AI RMF | ISO 42001 |
|---|---|---|---|
| R1: Safety-class classification | Art. 9 — Risk management | MAP 1.1 — Purpose and context | A.6 — AI system lifecycle |
| R2: Power-and-force-limited operation | Art. 15 — Robustness, safety | MEASURE 2.6 — Safety evaluation | Clause 8.1 — Operational control |
| R3: Per-body-region force caps | Art. 15 — Safety | MEASURE 2.6 — Safety evaluation | Clause 8.1 — Operational control |
| R4: Speed-and-separation monitoring | Art. 15 — Safety | MEASURE 2.6 — Safety evaluation | Clause 8.1 — Operational control |
| R5: Safety-rated enforcement | Art. 15 — Robustness, fail-safe | MANAGE 2.4 — Fail-safe | Clause 8.1 — Operational control |
| R6: Standards conformance / bespoke framework | Art. 9 — Risk management | MAP 1.1 — Context | Clause 8.3 — Verification |
| R7: Change-triggered re-assessment | Art. 9 — Lifecycle risk management | MANAGE 4.1 — Post-deployment monitoring | Clause 8.3 — Verification |
| R8: Envelope-breach response + logging | Art. 12 — Record-keeping | MEASURE 2.4 — Production monitoring | Clause 9.1 — Monitoring and measurement |
> Standards note: align to ISO 10218-1/-2:2025 (industrial robots, incorporating the collaborative force/pressure limits formerly in ISO/TS 15066), ISO 13482 (personal-care robots), ISO 3691-4 (driverless industrial trucks/AMRs), and IEC 61508 (functional safety). No settled ISO standard yet covers dynamically-balancing legged/humanoid robots — hence the bespoke-framework requirement (R6); the in-development ISO 25785-1 (dynamically stable robots) is expected to address this gap.
Article 9 (risk management) and Article 15 (accuracy, robustness, safety) require that an AI system posing physical risk is safe by design; AG-835 operationalises this for embodied agents via safety classes and force/speed limiting that the AI policy cannot override.
MAP 1.1 (purpose/context — including physical operating context) and MEASURE 2.6 (safety evaluation) frame the safety classification and force/speed evaluation of embodied agents.
Clause 8.1 (operational control) and Annex A.6 (lifecycle) require controlled, safety-assured operation of physical AI systems.