AxiSentinel extends continuous AI compliance evaluation from software to embodied systems: industrial robots, collaborative robots, autonomous mobile robots, and humanoid platforms operating in the physical world, not just in the cloud.
For most of the last two decades, robot safety was a mechanical-engineering discipline: force limits, e-stops, guarding. That changed the moment regulators started asking what happens when the robot's behavior comes from a model instead of a fixed program.
Every continuous AI-monitoring platform built so far was built for models that run in a data center. None of them were built to evaluate a robot that can share a floor with a person.
Physical AI sits at the intersection of machinery law and AI law. AxiSentinel tracks both, and builds RegDef packages the moment a standard is stable enough to evaluate against.
Industrial robot safety, now absorbing the former ISO/TS 15066 collaborative-robot thresholds directly into the main standard.
Replaces the 2006 Machinery Directive outright and, for the first time, explicitly brings AI-based safety functions and self-evolving code into scope.
Added 28 Jul 2026. A procurement and import restriction targeting specific foreign-made robot categories, not yet an operational compliance rule for deployers.
Successor to ISO 13482:2014, the baseline personal-care and service-robot safety standard, currently at Final Draft International Standard stage.
Formally titled "Collaborative Safety, Physical Contact with Robots, Part 1: Biomechanical Thresholds and Data" on ISO's work programme. A dedicated successor to the collaborative-robot force and pressure guidance now folded into ISO 10218, expanding it into its own standard.
General-purpose AI risk-management frameworks that already apply to any AI component inside a robotic system, physical embodiment or not.
From a single cobot on a factory floor to a fleet of humanoid platforms, AxiSentinel evaluates physical-AI behavior against whichever standard actually applies.
Nothing about AxiSentinel's core architecture changes for embodied systems. What changes is which RegDef packages are switched on and what telemetry the agents capture.
Agents capture robot and fleet telemetry on the cadence you configure, always-on or scheduled, in full rather than sampled, and never limited to a quarterly cycle.
Every telemetry event evaluated against the applicable physical-AI safety and machinery rule set, per AI event.
Every audit record is linked to the one before it in a signed, tamper-evident evidence chain, verifiable from the first event.
A compliance state change on one robot or an entire fleet propagates network-wide as it happens.
Nothing becomes a compliance finding until a qualified auditor reviews and signs it.
AXI-Node agents deploy on industrial control networks, mobile robot fleets, and fully air-gapped facilities through the .axibatch format, zero network connectivity required.
Tracks human involvement in autonomous physical actions and feeds it into AxiSentinel's oversight-gap scoring model.
The same agents generating compliance evidence watch for manipulated sensor inputs, spoofed perception data, and poisoned model updates before a re-certified release reaches production.
New machinery, robotics, or jurisdictional rules are added by encoding new RegDef packages. Deployed agents are never rebuilt.
Scoped to your organization during onboarding, not hard-coded into the platform.
A threshold breach becomes a flagged, timestamped, evidence-linked Provisional Alert, reviewed by a certified human auditor before anything counts as a finding.
The same architecture monitoring a warehouse robot fleet monitors a trading desk's model or a hospital's diagnostic AI. What changes is which RegDef packages are switched on.
Physical AI is scaling faster than the assurance industry built to evaluate it. AxiSentinel's architecture already does this for software AI; robotics is a new RegDef surface on the same platform, not a new product.
Whether it's a single collaborative robot or a fleet of humanoid platforms, AxiSentinel evaluates it the same way it evaluates any AI system: on the cadence you configure, always-on or scheduled, with full evidence, and with a human signature before anything counts as a finding.
