Path-planning claims age fast, so a serious AV company files the same idea more than once. Gatik AI did exactly that in 2022, with three grants — US11307594B2, US11320827B2 and US11487296B2 — all sharing the title "Method and system for deterministic trajectory selection based on uncertainty estimation for an autonomous agent."

The substance is in two words: deterministic and uncertainty. Classified under G05D 1/0221 (trajectory control) and G06N 20/00 (machine learning), the patents describe a vehicle that picks a path not just from what it perceives, but from how confident it is in that perception. When uncertainty is high, the deterministic selection biases toward caution — the car that is unsure slows and widens its margins.

“A system for deterministic trajectory selection based on uncertainty estimation includes a set of one or more computing systems.”— U.S. Patent No. 11,487,296 source

The lead patent's claim 1 spells out a dual-check architecture. The vehicle processes its inputs with a set of trained models to produce a first set of outputs carrying "a first set of uncertainty values," and separately processes the same inputs with an "out-of-distribution detection module" to produce a second set of uncertainty values. It then compares the first set against one threshold set and the second against another, and operates "according to the set of comparisons." Two independent doubts, two independent gates. That second module matters: per claim 7, its output is "a probability that the set of inputs is out of distribution relative to… training data" — it is not asking "how confident is the model," it is asking "have I ever seen anything like this before." Those are different failure modes, and the patent insists on testing both.

The fallback logic is the safety heart of the claim. Claim 2 says the first output includes a candidate trajectory; claim 3 spells out the disposition: if either uncertainty set exceeds its thresholds, "reject the first trajectory candidate," and only if both are within bounds does the vehicle maneuver along it. Claim 4 says a rejected candidate triggers a second trajectory "determined in absence of the set of trained models," and claim 5 names that source — "a classical motion planner," with claim 6 making it "a programmed fallback trajectory." That is the architecture in one breath: trust the learned planner when both confidence checks pass; when they don't, hand control to deterministic, non-learned code. The learned system is allowed to drive only when it can prove it isn't guessing.

The uncertainty model itself is textbook but precisely claimed. Claim 8 describes chaining trained models, the second evaluated on the first's outputs; claim 10 splits each model's uncertainty into "an epistemic uncertainty and an aleatoric uncertainty" — the standard distinction between uncertainty from limited knowledge (reducible with more data) and uncertainty from inherent randomness (irreducible). Claim 11 fuses the two per model. Claim 13 routes a "latent space representation" of the environment into the out-of-distribution module, and claim 14 says that module "comprises a generative adversarial network." A GAN trained to recognize in-distribution scenes will flag novelty when a live scene doesn't look like anything in its training set — a concrete mechanism behind the abstract phrase "out of distribution."

This is the honest engineering answer to a deep problem: a learned perception system does not just output "there is a pedestrian," it outputs a probability, and a driving policy that ignores the confidence is brittle. Tying trajectory choice to uncertainty estimation is how you keep the vehicle conservative exactly when it has the least reason to be aggressive. Claim 19 makes the trigger explicit — when "overall uncertainty exceed[s] a predetermined threshold," the system hands trajectory determination to "a programmed motion planner."

The "deterministic" half matters for a different reason: regulators and safety cases hate randomness. A trajectory selector that, given the same uncertainty, always makes the same choice is auditable in a way a stochastic one is not. The threshold-comparison structure of claim 1 is what delivers that determinism — identical inputs and uncertainty values yield identical accept/reject decisions, with no sampling step in the loop. Gatik's framing is built for the certification conversation, not just the technical one.

The fixed-route detail is where Gatik's specific business shows through the generic claim language. Claims 15–16 and the independent claim 20 describe a vehicle "configured for operation along a predetermined set of fixed routes," with the second uncertainty value tied to "a likelihood that the set of inputs is taken along a fixed route." Claim 16 is blunt: the out-of-distribution module "is designed to be overfit to data collected at the fixed route." For a general robotaxi, overfitting your novelty detector to specific roads would be a liability; for a middle-mile freight AV that runs the same lanes every day, it is exactly right — deviation from the known route is the anomaly worth braking for. The patent's safety logic and Gatik's fixed-route operating model are the same design choice viewed from two angles.

Filing three near-identical patents in one year is a strategy, not redundancy. It builds a thicket of overlapping claims around the company's core decision layer — the trajectory selector that is, for a middle-mile freight AV like Gatik's, the actual product. The repetition is defensive depth.

For readers tracking the AV field beyond the robotaxi headliners, Gatik is a reminder that fixed-route middle-mile freight is a real, narrower autonomy market — and that its IP fights are fought over exactly this uncertainty-aware decision logic, with its two-channel doubt, its classical fallback, and its deliberately overfit novelty detector, not over flashy perception demos.