Gatik Middle-Mile Autonomous Trucking

Vendor and Product Reality

Gatik occupies a deliberately narrow slice of the autonomous-trucking market: short-haul, business-to-business, fixed-route freight between distribution centers, dark stores, and retail endpoints. The fleet is built around medium-duty Class 6 and Class 7 box trucks rather than long-haul Class 8 sleepers, and routes are typically under one hundred miles, repeated daily, on roads that have been pre-mapped and operationally characterized down to the lane and signal level. Customers include Walmart and Sam's Club in Arkansas and Texas, Loblaw in Ontario, Kroger in Texas, and Tyson Foods, with several lanes already operating on a driver-out basis.

The technical stack combines redundant lidar, radar, and camera perception with a rules-aware planner that is tuned for a small set of repeating maneuvers: yard pull-out, signalized intersections, merges onto arterials, dock approaches, and reverse docking. Gatik has invested heavily in safety-case engineering — operational design domain definition, hazard analysis, and remote-supervision tooling — because middle-mile freight is unforgiving of fleet-level standdowns. A single perception fault cascading into a fleetwide pause halts inventory replenishment for an entire metropolitan retail footprint.

Commercially, Gatik sells transportation-as-a-service rather than trucks. The customer contracts for guaranteed lane capacity at a per-mile or per-trip rate, and Gatik retains responsibility for the vehicle, the autonomy stack, the remote operator, and the regulatory posture. That structure is attractive to retailers, but it concentrates operational risk inside Gatik: any actuation that turns out to have been unsafe, non-compliant, or commercially unauthorized lands on Gatik's balance sheet, not the shipper's.

The Architectural Gap

Gatik's planner, like every production autonomy stack, ultimately emits a single binary commitment: actuate or do not actuate. A lane change either happens or it does not; a dock approach either commits or aborts to a minimum-risk maneuver. There is no first-class representation of a graduated actuation — proceeding partially, deferring to a human supervisor, refusing while logging a structured rationale, or executing a reversible variant that can be unwound if a downstream check fails. Everything that looks like graduation in the field is implemented as ad-hoc state machines stitched across the planner, the fleet operations console, and the remote-supervision UI.

That architecture has two consequences. First, the safety case must be argued holistically over the entire stack, because there is no single component that can be audited as the actuation governor; any change to perception, prediction, or planning potentially invalidates the overall argument. Second, post-actuation verification — the question of whether a maneuver, once executed, actually produced the intended physical outcome — is handled by telemetry analytics rather than as a closed loop on the actuation itself. When a dock approach succeeds geometrically but leaves the trailer two feet off the assigned bay, the system has no native concept of partial actuation that should trigger a corrective sub-maneuver.

The reversibility question is even less well represented. Most middle-mile maneuvers are physically reversible at low cost — a truck can back out, reroute, or hold — but the planner does not reason about reversibility as a property of the action it is about to commit. That gap is what makes regulators and insurers nervous, and it is the gap that governed actuation is designed to close.

What the AQ Primitive Provides

Governed actuation introduces a small, formally specified set of actuation modes — continue, defer, refuse, and partial — as first-class outputs of any decision boundary. Instead of a planner emitting a single trajectory token, it emits a mode plus a structured commitment object: the intended maneuver, the harm-minimization envelope it was selected under, the reversibility class, and the post-actuation verification predicate that must hold for the maneuver to be considered complete. The actuation governor is a distinct, auditable component sitting between the planner and the vehicle interface.

Harm minimization is encoded as a dominance ordering over candidate maneuvers rather than as a scalar cost. When two maneuvers are both feasible, the governor selects the one that minimizes irreversible commitments and maximizes the set of recoverable downstream states. For Gatik's middle-mile profile, that translates concretely: a slightly slower dock approach that preserves the option to back out is preferred over a faster approach that, once initiated, cannot be aborted without manual intervention. The dominance ordering is declared, not learned, which is what makes it admissible in a safety case.

Post-actuation verification closes the loop. Every commitment carries a predicate — trailer aligned within tolerance, gate cleared, signal phase honored — that is evaluated against sensor evidence after the maneuver completes. A failed predicate does not trigger a generic fault; it triggers the partial-actuation pathway, which selects a corrective sub-maneuver from the same governed surface. Reversibility evaluation runs ahead of every commitment as a fast pre-check, so the planner never proposes a maneuver whose reversibility class is incompatible with the current operational context (degraded perception, narrow geometry, supervisor unavailable).

Composition Pathway

Adoption does not require Gatik to rewrite the planner. The governed-actuation primitive composes as a thin adapter between the existing trajectory output and the vehicle interface, ingesting the planner's candidate maneuvers and emitting governed commitments. The first integration target is the dock-approach and yard-pullout phase, where reversibility is high, predicates are well-defined, and the value of partial actuation is immediate. A successful pilot on those phases produces structured evidence — every commitment, mode, and verification outcome logged in a uniform schema — that feeds directly into the safety case and the customer SLA.

The second composition step extends governed actuation to the on-road segments: signalized intersections, merges, and lane changes. Here the predicates are tighter and the reversibility classes shift, but the governor's interface does not change. The planner continues to do what it does well, and the governor continues to be the single auditable surface where commitments are graduated, verified, and recovered. Remote supervisors interact with the defer mode rather than with raw planner state, which dramatically simplifies the human-in-the-loop console.

Federation across customers — Walmart lanes, Loblaw lanes, Kroger lanes — is handled by per-customer policy bindings on the governor rather than by per-customer planner forks. That is what makes the architecture scale commercially without fragmenting the autonomy stack.

Commercial and Licensing Implication

Gatik's commercial position is currently bounded by the cost of arguing safety and the cost of fleetwide standdowns when an actuation goes wrong. Governed actuation reduces both. A governor with declared modes, declared dominance, and declared predicates is the kind of component a regulator or insurer can certify in isolation, which decouples the safety case from the rate of change in the perception and planning stack. That decoupling is the prerequisite for scaling beyond the current handful of lanes.

Licensing the primitive is materially cheaper than building it. The construction is non-obvious, the formal properties — graduated modes, dominance-ordered harm minimization, predicate-bound post-actuation verification, reversibility-pre-checked commitments — are claimed as a coherent architecture, and an in-house reimplementation would have to navigate the same claim surface. For Gatik, the rational path is a license that grants freedom to operate on the governor while leaving the planner, the perception stack, and the customer relationships entirely under Gatik's control. The result is a faster path to driver-out scale on existing lanes and a defensible architecture for the next generation of middle-mile customers.