Physical Machine Safety
A VLM is fine until it picks the wrong path.
Foundation models are now planning robot motion. BoundaryAI is the deterministic safety layer between the planner and the actuator — at microsecond-class latency, inside the control loop.
What we enforce
Limits that don't ask the model nicely.
Speed & separation
ISO/TS 15066 envelopes evaluated per trajectory before motion.
Worker exclusion zones
Hard-blocked end-effector intrusion. Verified at the controller.
Force & torque caps
Commanded force clamped regardless of model output.
Mode supervision
State transitions gated by active permits and lockouts.
Tool & payload verification
End-effector and payload must match the active program before motion.
Telemetry & replay
Every refused command logged with full sensor context for incident review.
Performance
Microsecond-class decisions. Zero bypasses.
Measured 2026-09-16 on a self-hosted release build, x86 workstation, loopback, engine as the sole decision maker with no cloud in the path. Policy decision: 137–170µs median, 0.48–0.60ms p99, over 8,000 evaluations against ISO 10218-1 and ISO/TS 15066 rules. Full API round trip including audit write and signed transparency leaf: 4.8ms median. Sustained throughput 201 evaluations/sec per instance. Your controller is not our test client — figures scale with host and policy size. Methodology and raw percentiles available on request.
Bring AI to the floor without rewriting your safety case.
We integrate alongside your existing functional-safety stack — not in place of it.