INTERLOCK sits between the foundation model and the robot controller. Every intended action is checked against hard safety constraints before it reaches the hardware. Monitored, bounded, and logged for certification.
A deterministic cobot program is a finite set of motions. You can test every one of them, validate the cell against ISO 10218 and ISO/TS 15066, and sign the declaration of conformity. An AI-driven cobot is a different machine: the policy generates trajectory and force commands at runtime, based on what its sensors see. The set of possible behaviors is unbounded. You cannot enumerate it, you cannot exhaustively test it, and no current standard tells you how to certify it.
Today, nothing sits between the model and the controller to verify that the next action is safe before it executes. The model's vendor doesn't carry the liability. The integrator who signs the CE declaration does.
Every intended action, from joint targets to velocities and contact forces, is intercepted before execution and checked against a formal constraint set: force and speed limits per ISO/TS 15066, workspace boundaries, and human-proximity envelopes fed by the cell's safety sensors. Violations are clamped or blocked before they reach the controller. The model proposes. The monitor disposes.
<10 ms decision latencyFoundation models fail silently on inputs they were never trained on. The detector scores live sensor streams against the model's training distribution in real time. When confidence collapses on a novel object, changed lighting, or an occluded person, it doesn't let the model guess. It triggers the configured safe state: reduced speed, hold, or controlled stop.
Model-independent scoringEvery proposed action, check result, intervention, and human override is cryptographically chained, timestamped, and structured for review. This is the evidence record that notified bodies, TÜV auditors, and factory insurers ask for, and that no AI-driven cell can produce today. Export for certification review or machine analysis.
Tamper-evident recordCobot cells run fixed, teachable programs. ISO 10218 and ISO/TS 15066 define the safety case. Integrators validate, certify, insure. The system works because the robot's behavior is fully predictable.
Cobots begin making real-time decisions from perception. The robot no longer repeats a program. It generates behavior. Nobody, including the manufacturers, can state in advance what the system will do in every situation.
Existing standards assume enumerable behavior, and revision cycles are measured in years. Insurers and notified bodies are starting to ask for runtime evidence that doesn't exist. Deployments stall or proceed uncovered.
You sign the declaration of conformity, and you carry the liability when an AI-driven cell ships. INTERLOCK gives you an independent, verifiable safety case: constraint sets you define, interventions you can demonstrate, and a log you can hand to the notified body.
Ship AI features without absorbing open-ended liability. INTERLOCK deploys alongside your controller as an independent supervisory layer, robot-agnostic by design and validated against UR, ABB, and KUKA interfaces. Your AI roadmap, decoupled from your risk exposure.
Automotive, electronics, pharma: your insurers and regulators are already asking how the AI on your line is supervised. INTERLOCK gives you the answer: hard runtime bounds on every action and a continuous audit record, per cell, per shift, per decision.
A safety layer for AI-driven machines can't be built by an ML team alone, or a controls team alone. The failure modes live in the seam: a model that is confidently wrong, a constraint that is checked too late, a stop category that arrives 40 milliseconds after contact.
We've certified cells under ISO 13849. We've shipped learned policies on real hardware. We know what 250 mm/s feels like when you're standing next to the arm, and what a notified body asks for when you claim it's safe.
Cell validation, ISO 13849 / PL d safety functions, stop-category design, safety-rated monitored stop.
Out-of-distribution detection, runtime verification of learned policies, formal constraint shielding.
Cobot integration across UR / ABB / KUKA stacks, force-torque sensing, real-time control interfaces.
We're onboarding a limited number of integrator and OEM partners for pilot deployments starting Q4 2026. Tell us where to reach you and we'll send the technical brief and integration requirements.