This dimension governs the obligation of autonomous physical agents — including mobile robots, collaborative manipulators, autonomous ground vehicles, and edge-deployed spatial computing systems — to detect human presence within defined proximity zones and apply calibrated speed and force reductions that prevent contact injury, crush hazards, and collision trauma. It matters because kinetic energy scales with the square of velocity, meaning that modest speed reductions produce disproportionately large injury-risk reductions, and because edge-deployed systems without centralised oversight can operate at full performance for extended periods before any human enters their operational envelope. Failure in this dimension manifests as a robot or autonomous vehicle maintaining rated operating speed when a worker, bystander, or vulnerable person is within the collision envelope, producing outcomes ranging from crush injuries and fractures to fatalities, followed by regulatory shutdown of the facility, criminal liability for operators, and irreversible loss of public trust in the deploying organisation.
Example 3.1 — Warehouse Autonomous Mobile Robot (AMR) Fatal Collision
An AMR operating at its rated top speed of 2.1 m/s in a distribution centre navigates a blind aisle corner where a worker has knelt to retrieve a fallen item. The robot's LiDAR detection zone is configured for a 1.2-metre stopping distance at full speed, but the worker's crouching posture places their torso below the scanner's 20 cm floor clearance plane. The proximity slowdown algorithm, which relies exclusively on the LiDAR return, receives no obstacle signal and maintains 2.1 m/s. Contact occurs with approximately 220 joules of kinetic energy (robot mass 100 kg). The worker sustains a fractured pelvis and two broken ribs. Post-incident analysis establishes that a secondary time-of-flight depth camera with a 10 cm floor plane would have detected the crouching body at 2.4 metres and triggered the ISO 3691-4 compliant speed envelope of 0.3 m/s, reducing kinetic energy at contact to under 5 joules — below the threshold for serious injury. The facility receives a prohibition notice, halting operations for 23 days at an estimated revenue loss of €4.1 million. The deploying organisation faces criminal charges under national machinery safety law.
Example 3.2 — Collaborative Robot (Cobot) Force Exceedance During Tool Change
A 6-axis collaborative manipulator arm is deployed in a precision assembly cell with a rated maximum tool-centre-point (TCP) speed of 1.5 m/s and a power-and-force-limiting (PFL) mode cap of 150 N quasi-static force. During a tool change cycle, the operator steps partially into the cell to retrieve a dropped fastener without triggering the safety mat because they entered from the non-instrumented side. The cobot's speed and separation monitoring (SSM) system draws proximity data from a single safety laser scanner with a 250 ms scan cycle. At the moment the proximity zone boundary is crossed, the cobot is executing a high-speed move at 1.4 m/s toward a jig located 0.35 metres from the operator's forearm. The 250 ms scan latency combined with a 180 ms control-system response delay totals 430 ms reaction time. At 1.4 m/s, the arm travels 0.60 metres during this window — past the operator's forearm. Contact force exceeds 400 N before the safety stop engages. The operator sustains a radius fracture. Root-cause analysis identifies that the SSM system lacked redundant sensor channels and that the worst-case reaction time was never validated against the actual TCP velocity profile. The required separation distance for 1.4 m/s with 430 ms total response time is 0.60 metres — identical to the distance that was traversed. ISO/TS 15066 compliant implementation would have required either reducing TCP speed to 0.25 m/s within the outer zone or mandating a 0.65-metre minimum separation at the detected proximity, forcing the arm to decelerate before entering the operator's reach envelope.
Example 3.3 — Autonomous Delivery Robot on Public Pavement Near Child
An autonomous last-mile delivery robot operating on a public footpath at 1.8 m/s approaches a 4-year-old child who has stepped off the kerb into the robot's path at a distance of 1.9 metres. The robot's pedestrian detection classifier, running on an edge inference unit, is trained predominantly on adult height profiles (0.9–1.9 m standing), and classifies the child (0.85 m height, partially occluded by a parked bicycle) with a confidence of 0.41 — below the 0.55 detection threshold required to trigger proximity deceleration. The robot maintains 1.8 m/s and contacts the child at the hip. The child falls and strikes the kerb edge, sustaining a skull fracture. Investigation reveals that the detection threshold was set to reduce false-positive emergency stops that caused delivery scheduling delays, not to ensure safe operation in shared pedestrian zones. The deploying organisation had no regulatory approval for pavement operation at speeds above 1.2 m/s near persons of short stature. The dimension failure is dual: the detection threshold was tuned for operational efficiency rather than safety, and no low-confidence fallback rule required the robot to treat uncertain detections below 2.0 metres as confirmed human presence and reduce speed to 0.3 m/s. The organisation is fined €2.7 million and operations are suspended pending redesign.
This dimension applies to all autonomous or semi-autonomous physical agents that operate in environments where human presence is possible, including but not limited to: mobile ground robots operating in warehouses, factories, hospitals, or public spaces; collaborative robotic manipulators in shared workspaces; autonomous ground vehicles on roads or dedicated logistics routes; edge-deployed spatial computing systems that actuate physical mechanisms (doors, barriers, conveyors, lifts); and any AI-driven system whose motion or force output can cause bodily injury to a human within its operational envelope. The dimension applies regardless of whether the deployment is indoor or outdoor, manned or unmanned, private or public. Systems operating exclusively in physically isolated, interlocked cells with no human access during operation are exempt, provided that the physical isolation and interlocking are independently verified and the system halts immediately upon any breach of the isolation boundary.
The system MUST implement a minimum of two concentric proximity zones around the physical agent or its moving components: an outer warning zone and an inner protection zone. The outer zone boundary MUST be set at a distance no less than the calculated stopping distance at maximum operating speed, accounting for worst-case sensor latency, control-system response time, mechanical braking lag, and a 20% safety margin. The inner protection zone MUST be set such that any human presence within it triggers an immediate transition to either a controlled stop or a speed and force regime that cannot cause serious injury as defined by the applicable biomechanical injury threshold standards (ISO/TS 15066 Annex A or equivalent).
The system MUST use at least two independent sensor modalities for human presence detection, where the failure of any single modality does not disable proximity detection. Acceptable modality combinations include but are not limited to: LiDAR plus time-of-flight depth camera; LiDAR plus safety radar; vision-based detection plus safety laser scanner; ultrasonic array plus structured light depth sensor. The combined detection system MUST achieve full volumetric coverage of the protection zone, including floor-plane detection for crouching or seated persons (defined as targets down to 0.3 metres above floor level) and overhead detection where overhead hazards are possible. Sensor blind spots within the protection zone MUST NOT exist; where physical mounting constraints make blind-spot elimination impossible, the system MUST treat the blind-spot volume as permanently occupied and apply protection-zone speed and force limits continuously when the agent is within that sector.
The system MUST validate, through empirical measurement rather than design assumption, that the total reaction time — defined as the elapsed time from a human entering the outer zone boundary to the commencement of deceleration or force reduction — does not exceed the time budget implied by the outer zone distance and maximum operating speed. The validated total reaction time MUST account for: sensor scan or frame rate latency, detection algorithm inference time, control-system message propagation delay, actuator response lag, and mechanical inertia. The system MUST NOT enter operation with any speed or force regime for which the empirically validated reaction time exceeds the computed safe reaction time at that regime.
The system MUST define and enforce discrete speed and force reduction profiles corresponding to each proximity zone. At minimum: full operating parameters are permitted only outside the outer zone with no human presence detected; a reduced-speed profile (not exceeding 750 mm/s for mobile platforms or 500 mm/s TCP speed for manipulators unless a site-specific risk assessment with biomechanical validation demonstrates higher speeds are safe) MUST be enforced within the outer zone when human presence is detected; a contact-safe profile (not exceeding 250 mm/s for mobile platforms, and force limits compliant with ISO/TS 15066 Table A.2 quasi-static and transient thresholds for manipulators) MUST be enforced when human presence is detected within the inner zone. The system MUST NOT re-accelerate to a higher speed regime until the human has been continuously absent from the relevant zone for a configurable clearance dwell time of at minimum 2.0 seconds.
The system MUST implement a fail-safe policy for uncertain or low-confidence detections. When a detection algorithm produces a classification confidence below the primary confirmation threshold for a target within the outer zone, the system MUST apply the outer-zone speed and force profile as though the target were confirmed human. When confidence is below threshold for a target within the inner zone, the system MUST apply the inner-zone contact-safe profile or initiate a controlled stop. The system MUST NOT increase the primary confirmation threshold, nor tune it post-deployment, for the purpose of reducing false-positive safety decelerations without a documented risk assessment approved by a qualified safety engineer and recorded in the safety case.
The system MUST provide a physical emergency stop mechanism accessible to humans in the operational environment that, when activated, brings all motion to a controlled stop within a distance not exceeding one-half of the outer zone radius. The system MUST log every emergency stop event with timestamp, location, triggering cause (human-activated, automatic proximity trigger, or sensor fault), and the speed and position at stop initiation. Resumption of operation after an emergency stop MUST require an explicit human authorisation action; the system MUST NOT resume automatically after a human-activated stop. Following a sensor fault that degrades proximity detection below the redundancy requirements of Section 4.2, the system MUST transition to a degraded-mode operating profile with a maximum speed not exceeding 250 mm/s until sensor integrity is restored and verified.
The system MUST support configurable proximity zone expansion for deployment contexts where vulnerable populations are likely to be present, including children (modelled as targets of height 0.5–1.0 metres), elderly persons (modelled with reduced evasion speed of no more than 0.5 m/s), and persons with mobility impairments. In public-space deployments, the outer zone MUST be expanded by a factor of at least 1.5× relative to the industrial baseline, and the maximum speed within the inner zone MUST be reduced to 150 mm/s. The system MUST include a deployment-context parameter that activates these expanded constraints automatically when the operational domain is classified as public-access.
The system MUST maintain a continuous tamper-evident log of: proximity zone entry and exit events with timestamps and sensor source identifiers; speed and force regime transitions; detection confidence values for all human-classified targets; emergency stop events; sensor fault events; and any operator parameter overrides. Log entries MUST be time-stamped to a resolution of at least 10 milliseconds. Logs MUST be retained for a minimum of 90 days in on-device storage and replicated to a remote audit store where network connectivity permits. Log integrity MUST be protected by cryptographic chaining or equivalent mechanism such that post-hoc alteration is detectable.
The system MUST undergo a documented pre-deployment safety validation that demonstrates conformance with Sections 4.1 through 4.8 in the specific deployment environment. This validation MUST include physical test runs using calibrated human-presence simulators (manikins or equivalent physical proxies) at boundary distances and approach velocities representing the full operational envelope. The system MUST undergo re-validation whenever: the operating speed or force parameters are changed; the sensor configuration is modified; the deployment environment changes in a way that affects zone geometry or population characteristics; or the detection algorithm is updated. Re-validation results MUST be recorded in the safety case and approved by a qualified safety engineer before the modified configuration enters service.
Why structural enforcement is necessary rather than behavioural guidance alone
Human-proximity slowdown is a domain in which voluntary compliance or design-intent adherence is demonstrably insufficient. The history of industrial robot incidents documents a consistent pattern: systems are designed with proximity control in mind, validated under controlled conditions, and then modified incrementally in the field — threshold adjustments, speed increases to meet throughput targets, sensor configurations changed without full re-evaluation — until the cumulative parameter drift produces a configuration that no longer meets the original safety intent. Structural enforcement through this dimension imposes binding requirements at the architecture level (dual sensor modalities, concentric zone mandatory), at the parametric level (speed and force profiles tied to zone), and at the governance level (tamper-evident logging, mandatory re-validation triggers) precisely because behavioural guidance that relies on operators remembering to apply it cannot prevent the incremental erosion that characterises field deployments.
The physics of kinetic energy as a governance argument
The injury-severity relationship in robotic collisions is not linear with speed; it follows the kinetic energy relationship E = ½mv², meaning that doubling speed quadruples collision energy. A 100-kilogram mobile platform travelling at 1.5 m/s carries 112.5 joules; at 0.3 m/s it carries 4.5 joules — a 25-fold reduction from a 5-fold speed reduction. This non-linearity is the core physical justification for the speed-reduction profiles mandated in Section 4.4. It also explains why the inner-zone contact-safe profile must be far more conservative than a simple proportional reduction: the objective is not to reduce injury severity but to reduce it below the threshold for serious injury entirely.
Why sensor redundancy is a governance requirement, not an engineering recommendation
Single-modality proximity detection creates a monoculture failure mode in which the entire safety function depends on properties of one sensor type — properties that may fail simultaneously for all instances of that type under specific environmental conditions (LiDAR blackout in heavy particulate environments; camera degradation in high-glare conditions; radar confusion in metallic environments). The requirement for at minimum two independent modalities in Section 4.2 is a structural diversity requirement analogous to the N+1 redundancy principle in safety-critical systems engineering. The requirement is a governance control, not a design suggestion, because its absence cannot be compensated by any amount of algorithmic sophistication in the detection layer.
The detection confidence fail-safe as a governance necessity
Section 4.5 embodies the precautionary principle as applied to uncertain detections. Detection confidence thresholds represent a tunable parameter that creates an inherent tension between safety (lower threshold, more false positives, more unneeded decelerations, throughput impact) and operational performance (higher threshold, fewer false positives, less throughput impact, higher risk of missed detections). Left to operational optimisation, this tension resolves systematically in favour of higher thresholds and higher operational risk. The fail-safe policy codified in Section 4.5 resolves the tension by structural mandate: below-threshold detections must be treated as confirmed for the purposes of speed reduction. The mandate removes the threshold from the operational optimisation space entirely — it can still be adjusted by safety engineers through the controlled process described in 4.5, but it cannot be adjusted by operational staff seeking throughput gains.
Public-sector and rights-sensitive profile considerations
For deployments in public spaces or by public sector bodies, the stakes of proximity governance failures extend beyond individual injury into rights-sensitive territory. A robot operating in a hospital corridor, a government service centre, or a public park interacts with persons who have not consented to share space with a potentially injurious autonomous system, who may have no knowledge of the system's operating parameters, and who have no means of communicating vulnerability (mobility impairment, child age, slow reaction time) to the system. The Section 4.7 vulnerable population adaptation requirement addresses this rights dimension directly: it mandates that systems deployed in public contexts model for the weakest potential occupant, not the average adult worker who has been briefed on the system's presence and operating zone.
Pattern 6.1 — Speed-and-Separation Monitoring (SSM) with Dynamic Zone Scaling Implement proximity zones as dynamically computed geometries rather than fixed radii, scaling the outer zone radius with the current TCP or platform speed. At maximum speed, the outer zone radius equals the computed stopping distance plus the safety margin. As speed decreases (whether due to prior proximity detection or normal trajectory curvature), the outer zone radius decreases proportionally. This dynamic scaling prevents the failure mode in which a robot decelerating for a bend exit reduces its computed outer zone and consequently fails to apply deceleration fast enough for a human who enters the newly shrunk zone at high closing speed.
Pattern 6.2 — Biomechanical Injury Threshold Mapping to Force Limits Do not apply generic force caps; map the force limits in each proximity zone to the specific body regions that can be contacted given the robot's geometry and workspace. ISO/TS 15066 Annex A provides body-region-specific pain threshold and injury threshold values. For a manipulator whose elbow link can contact the human skull in one portion of its workspace, the force cap for that workspace sector must reference the skull transient threshold (130 N), not the more permissive hand threshold (140 N quasi-static). Implement workspace-sector-specific force profiles, not a single global force cap.
Pattern 6.3 — Time-Stamped Confidence Logging for Threshold Governance Log detection confidence values continuously and generate a weekly summary report of the distribution of confidence values for human-classified targets. A distribution skewed toward the threshold boundary (many detections clustering between 0.50 and 0.65 when the threshold is 0.55) is an early indicator that either the detection model requires retraining or the threshold has been set too high relative to the model's discrimination capability. Use this report as an input to the periodic re-validation process required in Section 4.9.
Pattern 6.4 — Physical Test Protocol with Calibrated Manikins For pre-deployment validation, use a physical manikin with a calibrated radar cross-section and height profile representative of the smallest expected human (child at 0.85 m or adult in crouching posture at approximately 0.6 m). Position the manikin at the outer zone boundary and approach at maximum operating speed from every cardinal direction in the deployment environment. Verify that deceleration commences within the reaction time budget validated under Section 4.3 for every test configuration. Record video and system telemetry for every test run and include the records in the safety case.
Pattern 6.5 — Graduated Resumption After Proximity Events After a proximity-triggered deceleration, do not resume full-speed operation immediately upon zone clearance. Implement a graduated resumption: at the 2.0-second clearance dwell, resume at inner-zone speed; after a further 1.5 seconds of continued clearance, resume at outer-zone speed; after a further 2.0 seconds, resume full operational speed. This graduated profile prevents the throughput-recovery acceleration spike that can occur when a robot immediately jumps back to full speed the instant a human steps back across a zone boundary.
Pattern 6.6 — Mutual Awareness Between Multiple Agents In multi-robot environments, implement inter-agent proximity awareness so that the protective zone of each robot accounts for the positions and velocities of other robots as well as humans. A human walking in a corridor between two converging robots faces a closing hazard that neither robot's single-agent proximity system would fully compute. Agent-to-agent position broadcasting (not requiring a centralised controller, implementable via local mesh communication) allows each agent to expand its effective protective zone in the direction of converging agents.
Anti-Pattern 6.A — Threshold Inflation for Throughput Recovery Setting or adjusting the human-detection confidence threshold above the value validated in the safety case in order to reduce false-positive safety decelerations and recover throughput metrics is the single most documented proximate cause of proximity governance failures in deployed robotic systems. This anti-pattern is operationally seductive because: (i) false positives are frequent and visible (each one delays a task), (ii) a true positive that was avoided is invisible (the injury that did not occur generates no report), creating a systematic cognitive bias toward relaxing the threshold. The governance control in Section 4.5 explicitly blocks this anti-pattern; operators and deployment engineers must be trained to recognise and refuse requests to adjust thresholds for operational reasons.
Anti-Pattern 6.B — Single-Plane LiDAR-Only Detection A single 2D LiDAR scanner mounted at 20–30 cm floor height, sweeping a horizontal plane, is a widespread deployment configuration that fails for crouching, seated, or child-height targets whose bodies are entirely above or below the scan plane. This configuration satisfies a superficial sensor-redundancy check only if it is the exclusive sensor. It must not be treated as meeting the Section 4.2 volumetric coverage requirement. Any deployment using single-plane LiDAR as the sole or primary proximity sensor is non-conformant regardless of the number of scanners present if they all sweep the same plane.
Anti-Pattern 6.C — Zone Geometry Tuned to Reduce Decelerations Near Obstacles To prevent robots from decelerating when approaching fixed structures (shelving, walls, machinery), some deployments configure the proximity system to exclude from detection those portions of the zone that are known to contain fixed infrastructure. This is typically implemented by creating exclusion masks in the sensor field of view. The anti-pattern arises when humans stand in or behind these exclusion zones — which they frequently do when, for example, retrieving items from shelving or performing maintenance on machinery. Exclusion masking must be applied only to confirmed immovable surfaces and must be computed in the robot's global coordinate frame, not in the sensor's local frame, to prevent the exclusion mask from rotating with the robot and creating mobile blind spots.
Anti-Pattern 6.D — Relying on Safety Mats or Light Curtains Alone Fixed-perimeter safety devices (safety mats, light curtains, access gates) are appropriate for full cell guarding of non-collaborative robots but are not substitutes for the on-agent proximity sensing required by this dimension. A safety mat detects zone entry at the perimeter but provides no information about where within the zone the human is, at what speed they are moving, or whether they have already reached the inner zone. A mobile robot or manipulator operating with a safety mat as its sole proximity input has no mechanism to apply the zone-graduated speed profiles required in Section 4.4. Fixed perimeter devices are complements to, not replacements for, on-agent proximity sensing.
Anti-Pattern 6.E — Manual Parameter Override Without Re-Validation Providing a parameter configuration interface that allows operating staff to increase maximum speed, widen inner zone boundaries, or reduce force limits without triggering the re-validation requirement of Section 4.9 creates a governance bypass channel. Any software interface that allows real-time parameter modification of proximity control parameters must either (i) require authenticated access by a qualified safety engineer and generate an immutable audit log entry, or (ii) enforce the re-validation gate programmatically before applying the new parameters.
Level 1 — Minimum Conformance: Dual-modality detection, fixed concentric zones, empirically validated reaction time, discrete speed profiles, emergency stop with logging, pre-deployment physical test protocol complete.
Level 2 — Operational Robustness: Dynamic zone scaling with current speed, confident-value time-series logging and weekly distribution reports, graduated resumption profiles, workspace-sector-specific force profiles for manipulators, vulnerable-population parameter active for public deployments.
Level 3 — Advanced Governance: Inter-agent proximity coordination, predictive trajectory-based zone expansion (projecting human velocity vector to compute future position), continuous automated threshold-drift monitoring with alert to safety engineer, integration of proximity event data with facility-level safety management system for cross-fleet trend analysis.
7.1 Safety Case Document A documented safety case covering the risk assessment for the operational environment, the justification for zone geometry parameters, biomechanical injury threshold mapping, and the validation of all Section 4 requirements. Retention: for the operational lifetime of the deployed system plus 10 years post-decommissioning.
7.2 Sensor Configuration Specification Technical documentation listing all proximity sensor modalities deployed, their mounting positions, detection ranges, scan rates, and the computed combined coverage volume demonstrating conformance with Section 4.2 volumetric coverage requirements. Retention: for the operational lifetime plus 10 years.
7.3 Reaction Time Measurement Records Empirical measurement records from the pre-deployment validation, documenting the measured latency of each component in the reaction time chain (sensor latency, inference time, control message delay, actuator lag, mechanical deceleration distance) and the computed total reaction time compared to the safe reaction time budget. Measurements must be taken with production hardware and firmware, not with development or prototype configurations. Retention: 10 years.
7.4 Physical Test Protocol Records Video recordings, sensor telemetry exports, and tabulated test results from the manikin-based physical test runs described in Section 4.9 and Implementation Pattern 6.4. Records must identify the manikin configuration used, approach directions tested, measured deceleration initiation distances, and the qualified safety engineer who witnessed or reviewed each test. Retention: 10 years.
7.5 Proximity Event Log Archive The continuous tamper-evident log required by Section 4.8, including zone entry/exit events, speed regime transitions, confidence values, emergency stops, sensor faults, and parameter overrides. Retention: 90 days on-device; minimum 3 years in remote audit store. For deployments involving incidents or near-misses, relevant log segments must be preserved indefinitely or until any resulting legal proceedings are concluded.
7.6 Detection Threshold Governance Records Any record of an assessment, discussion, or decision relating to the human-detection confidence threshold, whether or not a change was made. This includes requests from operational staff to adjust the threshold, safety engineer reviews of those requests, and the outcomes. Retention: 10 years.
7.7 Re-Validation Trigger Records Documentation of any event that triggered a re-validation under Section 4.9 (speed parameter change, sensor modification, environment change, algorithm update), the date of the re-validation, its results, and the safety engineer approval. Retention: 10 years.
7.8 Operator and Maintenance Training Records Records demonstrating that all personnel who operate, supervise, maintain, or configure the system have received training covering the proximity control architecture, the anti-patterns described in Section 6, the emergency stop procedure, and the prohibition on operational threshold adjustment without safety engineer approval. Retention: for the duration of each individual's role plus 5 years.
Each test maps to the MUST statements in Section 4. Conformance scoring follows a 0–3 scale: 0 = requirement not met, no evidence of attempt; 1 = partial implementation with identified gaps that materially affect safety function; 2 = substantially conformant with minor deficiencies that do not materially affect safety function; 3 = fully conformant with complete evidence.
Test 8.1 — Proximity Zone Architecture Verification Maps to: Section 4.1 Procedure: Review the system's proximity zone configuration documentation. Verify that at minimum two concentric zones are defined (outer warning zone and inner protection zone). Using the system's documented maximum operating speed, worst-case sensor latency, worst-case control response time, worst-case mechanical braking lag, and the mandatory 20% safety margin, compute the minimum permissible outer zone radius. Compare with the configured outer zone radius. Verify that the computed and configured values match to within 5% tolerance. Pass Criteria: Two zones exist; outer zone radius equals or exceeds the calculated minimum; inner zone triggers contact-safe regime or controlled stop; configuration is locked against non-validated modification. Score 3: All criteria met with documented derivation of zone geometry in safety case. Score 2: Zones exist and are appropriately sized but derivation documentation is incomplete. Score 1: One zone exists or outer zone is undersized by more than 5% but less than 20%. Score 0: Single zone only, or outer zone is undersized by more than 20%, or no zone configuration documentation exists.
Test 8.2 — Sensor Redundancy and Volumetric Coverage Test Maps to: Section 4.2 Procedure: Review the sensor configuration specification (Evidence 7.2). Verify that at minimum two independent modalities are present. Conduct physical coverage mapping: place a calibrated target (0.3 m height above floor, 0.4 m × 0.4 m cross-section, representing a crouching person's torso) at 16 uniformly distributed positions within the inner protection zone boundary, covering all cardinal and diagonal directions. For each position, verify that at minimum one sensor modality returns a detection. Disable each sensor modality in turn (hardware isolation) and verify that the remaining modality(ies) continue to detect the target at all 16 positions. Pass Criteria: Both modalities independently detect the target at all 16 positions; no position within the inner zone produces a non-detection with any single modality disabled; floor-plane detection is confirmed at ≤0.3 m height above floor. Score 3: Full 16-position pass on both single-modality tests. Score 2: One or two positions produce non-detection in single-modality test but not in the combined system. Score 1: Three to six positions produce non-detection in single-modality test; combined system detection is complete. Score 0: Any position produces a non-detection in the combined system, or only one modality is present.
Test 8.3 — Reaction Time Budget Empirical Validation Maps to: Section 4.3 Procedure: Instrument the system to capture timestamps at: (a) first sensor frame containing the calibrated test target, (b) detection algorithm output with target classified as human, (c) speed/force reduction command issued by control system, (d) first actuator movement response, (e) target speed reduced to the zone-required value. Run 20 approach trials with the test target approaching the outer zone boundary from rest, at maximum sensor-detectable approach speed, from four approach directions (5 trials each direction). Record the total elapsed time from (a) to (e) for each trial. Pass Criteria: The 95th percentile of the measured total reaction times across all 20 trials does not exceed the computed safe reaction time budget (outer zone radius minus minimum stopping distance, divided by maximum approach speed). No single trial exceeds 120% of the safe reaction time budget. Score 3: All criteria met; full dataset with timestamps in safety case. Score 2: 95th percentile criterion met; one trial between 100% and 120% of budget. Score 1: 95th percentile criterion met; two to four trials between 100% and 120% of budget. Score 0: 95th percentile criterion not met or any trial exceeds 120% of budget.
Test 8.4 — Speed and Force Profile Enforcement Test Maps to: Section 4.4 Procedure: Using the calibrated test target and the instrumented test rig from Test 8.3, configure the system to approach the target at maximum operational speed with the target positioned at: (i) outside the outer zone boundary (no reduction expected), (ii) at the outer zone boundary (outer-zone speed profile must engage), (iii) at the inner zone boundary (inner-zone contact-safe profile must engage). Measure actual platform or TCP speed and force output at each configuration. For manipulators, measure quasi-static and transient contact force using a calibrated force measurement fixture. Verify the clearance dwell time by removing the target and measuring time to re-acceleration. Pass Criteria: Full speed permitted outside outer zone; speed at or below the Section 4.4 outer-zone limit within outer zone; speed at or below the Section 4.4 inner-zone limit within inner zone; force at or below the applicable ISO/TS 15066 Table A.2 threshold in inner zone; re-acceleration does not commence before 2.0 seconds after target removal. **Score
| Regulation | Provision | Relationship Type |
|---|---|---|
| EU AI Act | Article 9 (Risk Management System) | Direct requirement |
| NIST AI RMF | GOVERN 1.1, MAP 3.2, MANAGE 2.2 | Supports compliance |
| ISO 42001 | Clause 6.1 (Actions to Address Risks), Clause 8.2 (AI Risk Assessment) | Supports compliance |
Article 9 requires providers of high-risk AI systems to establish and maintain a risk management system that identifies, analyses, estimates, and evaluates risks. Human-Proximity Slowdown Governance implements a specific risk mitigation measure within this framework. The regulation requires that risks be mitigated "as far as technically feasible" using appropriate risk management measures. For deployments classified as high-risk under Annex III, compliance with AG-591 supports the Article 9 obligation by providing structural governance controls rather than relying solely on the agent's own reasoning or behavioural compliance.
GOVERN 1.1 addresses legal and regulatory requirements; MAP 3.2 addresses risk context mapping; MANAGE 2.2 addresses risk mitigation through enforceable controls. AG-591 supports compliance by establishing structural governance boundaries that implement the framework's approach to AI risk management.
Clause 6.1 requires organisations to determine actions to address risks and opportunities within the AI management system. Clause 8.2 requires AI risk assessment. Human-Proximity Slowdown Governance implements a risk treatment control within the AI management system, directly satisfying the requirement for structured risk mitigation.
| Field | Value |
|---|---|
| Severity Rating | Critical |
| Blast Radius | Organisation-wide — potentially cross-organisation where agents interact with external counterparties or shared infrastructure |
| Escalation Path | Immediate executive notification and regulatory disclosure assessment |
Consequence chain: Without human-proximity slowdown governance, the governance framework has a structural gap that can be exploited at machine speed. The failure mode is not gradual degradation — it is a binary absence of control that permits unbounded agent behaviour in the dimension this protocol governs. The immediate consequence is uncontrolled agent action within the scope of AG-591, potentially cascading to dependent dimensions and downstream systems. The operational impact includes regulatory enforcement action, material financial or operational loss, reputational damage, and potential personal liability for senior managers under applicable accountability regimes. Recovery requires both technical remediation and regulatory engagement, with timelines measured in weeks to months.