AV actuator safety gate · kinematic invariants

TransitOS

A deterministic safety barrier in pure Rust between the neural driving planner and the drive-by-wire actuators. No command that violates a kinematic envelope reaches a steering or braking motor.

Actuator gate telemetryenvelopes 100% enforced · 14 hallucinations intercepted
braking lateral requested +34° steering clamped +4.2° Kamm friction limit μ the requested vector leaves the circle; the substitute sits on it
stopping distance against obstacle clearance, worst-case adhesion braking distance under worst-case μ obstacle, radar and LiDAR range clearance > 0, proved TTC invariant: stopping distance < clearance, always
Gate latency≤ 0.42 mshard real-time
Braking decel bound≤ 0.85 gproved
Hallucinations blocked14intercepted
Planner trustZerountrusted caller

01 — the gateThe planner is an untrusted caller

An end-to-end neural driving policy is a black box that can be fooled by pixels. Treating it as anything else is what puts a statistical safety argument in the path of a deterministic physical limit.

TransitOS interposes itself between the planner and the actuator bus and treats every command as untrusted input. When the planner asks for a hard turn at speed, the gate evaluates it against Kamm's friction circle: longitudinal braking and lateral acceleration together cannot exceed what the tyres can hold. A request outside the circle is not softened or averaged — it is rejected, and the nearest verified-safe command is dispatched in its place.

[AI-INPUT]End-to-end vision model commanded +34° steering at 90 km/h
[KINEMATIC-FILTER]Lateral rollover threshold exceeded: REJECTED
[SAFE-SUBSTITUTE]Clamped to maximal stable turn rate: +4.2°
[ACTUATOR-BUS]CAN FD verified command dispatched safely

02 — the invariantsWhat is proved, and what happens when the planner dies

The friction circle is enforced continuously: combined lateral and longitudinal demand stays inside the tyre limit, which is what keeps a commanded turn from becoming a loss of control.

The time-to-collision invariant is a stopping-distance proof rather than a warning threshold. Under worst-case road adhesion, the distance required to stop is shown to remain strictly less than the obstacle clearance measured by radar and LiDAR. If that cannot be shown, the command does not pass.

Deterministic fallback. When the planner crashes, desynchronises or times out, the gate takes control in under a millisecond and brings the vehicle to a minimum-risk stop. The failure mode of the planner is not the failure mode of the vehicle.

03 — comparedAgainst a monolithic AI stack

CharacteristicDirect end-to-end AI stackTransitOS actuator safety gate
safety argumentstatistical probability over empirical milesdeterministic mathematical proof
adversarial inputsusceptible to optical illusion attacksphysics invariants cannot be fooled by pixels
execution predictabilityvariable GPU inference, 20–80 msdeterministic hard real-time, ≤ 0.5 ms
certification pathnear-impossible ISO 26262 ASIL D audita formally proven safe subset, amenable to ASIL D
git clone https://github.com/KELLERBABG/TransitOS.git
cd TransitOS && cargo test        # the safety gate test suite
TransitOS · Keller Systems repository  /  kellersystems.dev