The word "interpretable" in a motion-planner claim is doing safety-and-regulatory work. UATC's grant US11755018B2 ("End-to-end interpretable motion planner for autonomous vehicles," issued September 12, 2023; inventors Wenyuan Zeng, Wenjie Luo, Abbas Sadat, Bin Yang, Raquel Urtasun) claims a planner trained end-to-end — perception to trajectory in one learned pipeline — that nonetheless produces intermediate representations a human can inspect. The interpretability is the inventive hedge against the black-box problem.

Read claim 1 closely and the "interpretable" label resolves into a concrete data structure. The recited machine-learned model takes sensor data and map data and generates a cost volume — a structure that "indexes, at each of a plurality of timesteps, cost data associated with a plurality of locations in the environment." That cost volume is the load-bearing intermediate. Rather than a single network emitting a steering command from raw pixels, the claim forces the model to first lay down a spatiotemporal cost field over the scene, then choose a trajectory against it. The field is something an engineer can render and audit; it is the artifact that makes the planner explainable.

The trajectory-selection step is equally specific. Claim 1 requires selecting a target trajectory "from a plurality of potential trajectories based on a comparison" of those candidates, where for each candidate — described by waypoints tied to the plurality of timesteps — the model uses "the waypoints and the plurality of timesteps to obtain cost data from the cost volume for determining a respective trajectory score." This is a sampling-and-scoring planner, not a regression that hallucinates a path: a set of candidate trajectories is laid over the learned cost field, each waypoint is looked up at its matching timestep, and the costs are aggregated into a score. Claim 14 spells out the lookup mechanism — applying "a respective filter to the cost volume" at each waypoint's timestep and combining the resulting values — while claim 20 recites that the cost volume holds "precomputed scores," which is what lets the scoring run cheaply at control rates.

“Systems and methods for generating motion plans including target trajectories for autonomous vehicles are provided.”— U.S. Patent No. 11,755,018 source

The dependent claims are where "interpretable" becomes enforceable rather than aspirational. Claim 2 adds first convolutional layers that "generate one or more intermediate representations associated with at least one of an object detection or an objection prediction," and claim 3 pins those representations down as "one or more bounding boxes" plus "one or more motion predictions." Claim 11 then makes the inspection explicit: the intermediate representations are exposed as a second output of the model, alongside the trajectory. So the planner does not merely use detections and predictions internally — it surfaces them, giving a safety reviewer the perception artifacts (what the car saw, where it thought things were going) that sit behind the chosen path. That is the difference between a black box and a model that shows its work.

The training story closes the loop. Claim 5 recites that the detection layers "are optimized based on an output of the" cost-volume layers, claim 6 adds "multitask training with supervision for perception and motion planning," and claim 7 a "total loss function" with both a perception loss and a motion-planning loss. Claim 8 decomposes the perception loss into a classification loss ("distinguishing a vehicle from a background") and a regression loss ("generating object bounding boxes"), and claim 9 grounds the planning loss in "one or more human-driven trajectories." The model is end-to-end because a single joint loss shapes both the intermediates and the final trajectory; it is interpretable because the perception half of that loss forces the intermediates to mean something a human recognizes. Those two properties — usually in tension — are reconciled by the multitask objective. CPC-wise, G05D 1/0212 (position control toward a target) and G01C 21/3453 (route computation) mark the planning-and-navigation core.

For the control beat, this is where engineering meets the regulatory frame. A purely end-to-end planner is hard to certify because you cannot say why it did what it did; an interpretable one gives you the artifact a safety case needs — here, an auditable cost volume plus exposed detections and motion forecasts. Fencing the interpretable-end-to-end combination is fencing the version of learned planning that can actually survive a safety review, which is the version that ships. Claim 13's vehicle-control-signal step makes the deployment intent explicit: the scored trajectory becomes actual control commands.

From a portfolio angle, this is another Urtasun-group asset (the same inventor cluster behind UATC's 3D-detection family), and it fits a pattern: fence the learned-autonomy primitives that bridge research and deployment. The independent media claims (16 and 18) re-fence the same cost-volume-and-scoring method as software and as a method, so the enforceable surface is the union of the apparatus, media, and method forms — not the device alone.

Caveats. End-to-end and interpretable planning each have substantial prior art; the grant survives on the specific architecture — the indexed cost volume, the waypoint-and-timestep lookup, the exposed intermediate output — not on the aspiration. "Interpretable" is doing real work only because claim 2 and claim 11 recite the mechanism that produces and surfaces the intermediates. Read the independent claims for exactly which representation is exposed and how the score is computed.

For the file: a deployment-flavored learned-planning grant whose value is the interpretability limitation, made concrete by an indexed cost volume and an exposed detection/prediction output. Pull US11755018B2 in the patent record, read claims 1, 11, and 14 for the recited interpretable intermediate and the scoring lookup, and file it under planning IP built to clear a safety case.