A single-code CPC is the patent at its most concentrated. NVIDIA's grant US12496714B2 ("Collision-free motion generation," issued December 16, 2025; inventors include Balakumar Sundaralingam, Caelan Reed Garrett, Nathan Donald Ratliff, Karl Van Wyk, Dieter Fox) classifies into exactly one bucket: B25J 9/1664 — manipulator control based on a sensed or known environment. The whole claim is about generating motion that doesn't hit anything.

The inventor list is the tell — this is the team behind NVIDIA's GPU-accelerated motion-planning work (the cuRobo line) inside the Isaac robotics stack — and claim 1 reads exactly like that lineage. It is a three-stage parallel pipeline. First, "at least one first parallel processing unit (PPU)" generates "a plurality of seed trajectories in parallel" from an initial joint configuration and a set of "seed joint configurations that would position at least a portion of a first manipulator in a goal pose." Second, "at least one second PPU" modifies "the plurality of seed trajectories in parallel to obtain a plurality of modified trajectories." Third, the method selects one "based at least in part on cost values calculated for the plurality of modified trajectories" and provides the resulting motion plan to a manipulator to execute. The word "parallel" appears at every stage — that is the inventive substance, and it is what turns a classically slow planning problem into something that runs at control rates on a GPU.

The dependent claims name the actual optimization recipe, which is unusually specific for a motion-planning claim. Claim 2 makes the modification step iterative "until at least one stopping condition is satisfied." Claim 3 requires "a gradient-free optimization, and a gradient-based optimization" — both — and claims 4 and 5 name them: a "particle-based optimization" for the gradient-free stage and, for the gradient-based stage, "a Limited-memory Broyden-Fletcher-Goldfarb-Shanno (L-BFGS) algorithm." That hybrid — a particle method to explore broadly, then L-BFGS to refine — is a concrete, recited algorithm, not a generic "optimize the trajectory." Claims 6 and 7 give the stopping conditions (task cost below "a threshold value," or "a desired number of iterations").

“Apparatuses, systems, and techniques to perform collision-free motion generation (e.g., to operate a real-world or virtual robot). In at least one embodiment, at least a portion of the collision-free motion generation is performed in parallel.”— U.S. Patent No. 12,496,714 source

The seed-generation stage is just as detailed. Claim 8 builds the seed trajectories by "using inverse kinematics to obtain the plurality of seed joint configurations," modifying them on "at least one third PPU," then "interpolating between the initial joint configurations and the plurality of modified joint configurations." Claim 10 offers an alternative seeding path via "geometric planning," and claim 22 fleshes that out as a parallel graph search — "searching along directions between at least one vertex of a graph and a plurality of target points…in parallel for at least one endpoint that does not collide with any obstacles," then adding the vertex and edge. So even the planning-graph expansion is parallelized. This is the GPU thesis applied end to end: seed in parallel, optimize in parallel, search the roadmap in parallel.

The collision math — the "collision-free" in the title — is recited too. Claims 24–28 define the collision cost: claim 25 walks a virtual object between positions and accumulates cost when it "would collide with at least one second virtual obstacle," claim 27 handles self-collision by treating the robot as "pairs of smaller objects" and "calculating distances between any of the pairs…in parallel," and claim 28 uses parallel "forward kinematic processes" to locate links and "backward kinematic processes" to obtain "gradients" for the optimizer. That last detail is the bridge: differentiable, parallel kinematics is what feeds the L-BFGS gradient step. Claims 12–14 add time-scaling under "velocity," "acceleration, or…jerk" constraints, so the output trajectory is dynamically feasible, not just collision-free in geometry.

For the manipulation beat, collision-free motion generation is a foundational primitive every manipulator needs, and the single tight B25J code signals a focused claim on a specific generation method. The inventive content is plainly the parallel/GPU formulation — the hybrid particle-plus-L-BFGS optimizer over GPU-parallel differentiable collision and kinematics — which is exactly NVIDIA's structural advantage. Claims 18 and 17 extend the manipulator to "an autonomous vehicle, an aerial drone, a cleaning device, a legged robot, or a walking robot," keeping the platform language wide.

From a portfolio angle, this is NVIDIA fencing a core Isaac motion-planning primitive, complementing its reactive-interaction grant (US12454050B2, 2025) and its robotic-control-system grant (US12194632B2, 2025). The pattern is consistent with its autonomy strategy: own the foundational robotics-software primitives that every robot builder running on NVIDIA hardware will use, so the moat is the software layer of the platform. The independent system (claim 15) and PPU (claim 29) forms re-fence the same parallel pipeline as hardware.

Caveats. Collision-free motion planning has one of the deepest prior-art bases in all of robotics; a broad claim is unthinkable, so the grant lives on the specifics — the parallel seed/optimize/select pipeline and the named particle-plus-L-BFGS hybrid. An accused system likely has to do the modification in parallel and select on cost to read on claim 1. Read the independent claims with claims 3–5 and 22 for the optimizer and the parallel graph search.

For the file: a focused, GPU-flavored collision-free motion-generation grant — cuRobo-style — inside NVIDIA's Isaac primitive portfolio. Pull US12496714B2 with the sibling 2025 NVIDIA robotics grants in the patent record; read claim 1 plus claims 3–5 and 28 for the parallel hybrid optimizer and differentiable collision/kinematics — that parallel formulation is the asset.