Is the robot fleet actually funded? It is only funded if it is fast enough to pay back — and safety is the governor on speed. NVIDIA's June 9, 2026 grant US12649232B2, "Speed determination in robotics systems and applications," claims methods for a robot to determine how fast it can move. Its CPC tags, B25J 9/1651 and B25J 9/1666 (manipulator motion planning) with B25J 13/089 (sensing-based control), place it in the heart of robot motion control.

Translate that into a cost line. A warehouse or factory robot that must slow to a crawl near humans moves less product per hour; one that can compute a safe higher speed moves more. Throughput is the numerator of every robotics ROI calculation, and safe-speed determination sets it. A patent on doing that determination well is a patent on the lever between "deployable" and "deployable and profitable."

“In various examples, a technique for generating speed change decisions for a mobile robot includes identifying, using one or more maps of a physical environment, one or more obstacles associated with one or more portions of a path of the mobile robot in the physical environment.”— U.S. Patent No. 12,649,232 source

The claims show the method is more sophisticated than a proximity slowdown, and that sophistication is the economic point. Independent claim 1 generates "one or more speed constraints, each speed constraint specifying a speed limit for a respective portion of the path," then searches for a speed profile by building "a representation of a path time-space region having a spatial dimension based at least on a length of the path and a time dimension based at least on a time of travel along the path." The robot does not merely brake when something is close; it plans a speed profile through a map of where obstacles will be and when, then searches that space-time region for "a particular speed profile that does not intersect the time-space representations of the speed constraints." This is optimization, not reflex — the difference between a robot that stops dead and one that threads a corridor at the fastest speed that is still provably safe.

The dependent claims tie that math back to the warehouse floor. Claim 4 lists the obstacle types it reasons about — "an automatic door, stairs, or a pedestrian" — the exact furniture of a fulfillment center. Claim 6 handles "a dynamic obstacle... wherein at least a portion of the dynamic obstacle moves over time," so the robot's speed limit depends on where a moving person will be at a given instant, not just where they are now. That distinction between static and dynamic constraints (claim 12 splits them explicitly) is where the throughput is won: a robot that can keep moving around a predictably-moving human, instead of freezing, is a robot with a higher effective duty cycle.

The output side of the claims is where "decision" becomes "control signal," and where the economics get concrete. Claim 8 has each speed change decision specify "a speed-related action to be performed," and claim 9 narrows those actions to "a stop action or an accelerate action" — the robot is not just told a speed limit, it is handed a discrete instruction to brake or to speed up. Claim 11 then translates decisions into "one or more speed directives, each... specifying a distance range on the path, a time range, and a speed limit that applies to the mobile robot in the distance range during the time range." That is a fully operational control schema: where, when, and how fast. A patent that runs from "identify obstacles on a map" all the way to "accelerate here, between these points, for this window" owns the entire chain from perception to throttle — which is precisely the chain that determines how many units per hour a deployed robot clears.

This is also the NVIDIA platform pattern showing up in robotics, not just driving. The company is not selling a robot; it is selling the motion-and-perception intelligence that other people's robots run on. Claim 17's deployment list says so directly — the processors may sit in "a control system for an autonomous or semi-autonomous machine," a "system implemented using a robot," one running "large language models," or one "implemented at least partially in a data center" or "using cloud computing resources." A safe-speed method written to run on everything from an edge robot to a data-center planner is one more piece of the stack, and one more reason a robotics OEM building on NVIDIA's platform stays on it.

For a capex-minded reader, the metric to watch is throughput per deployed robot over successive generations. If it rises, part of the gain is better safe-speed control — planning a profile through space-time rather than braking on proximity — and that control increasingly comes from platform suppliers' IP rather than the OEM's own work. The productivity curve and the supplier's patent estate move together. The grant runs to 19 claims across a method, a processor, and a system form, the breadth that lets the holder assert it against an OEM, a chip, and an integrated system alike.

The honest limit: a speed-determination grant is a method, not a benchmark. The 19 claims describe how to compute a safe speed profile; they do not tell you how much throughput it unlocks in a given building or who licenses it, and a granted claim is not an enforced one. It tells you NVIDIA is patenting the safety-throughput tradeoff — the exact constraint that decides whether a commercial robot earns its keep.

The takeaway for the money desk is that in robotics, safety is not a compliance footnote — it is the unit-cost driver. The fastest a machine can safely move is the ceiling on what it can earn, and this patent claims a way to push that ceiling higher by reasoning about moving obstacles in space and time. NVIDIA patenting how that ceiling is computed is the company quietly positioning itself at the most economically load-bearing point in the robot.