Where a patent gets classified is the first thing it tells you about its scope. UBTECH's grant US10821605B2 ("Robot motion path planning method, apparatus and terminal device," issued November 3, 2020; inventors Youjun Xiong, Jinghua Tang) sits squarely in B25J 9/1666 — the CPC bucket for programme-controlled manipulators that plan a path. That single classification quietly excludes the whole G05D mobile-robot world. This claim is about an arm, not a rover.

And claim 1, once you read it, is far more specific than the broad title implies — it recites a particular refinement algorithm, not "path planning" in general. The method starts by "planning a planned path for a robot in a current scene using an open motion planning library (OMPL) database" — OMPL being the well-known sampling-based planning library, so the initial path comes from an off-the-shelf RRT/PRM-style planner. It then "sets a shortest ideal path as an initial ideal path" (the straight-line "shortest path between the start point and the end point") and — here is the inventive core — "calculating a new path between the planned path and the initial ideal path using a dichotomy method." Dichotomy is bisection: the algorithm repeatedly halves the gap between the safe-but-wandering sampled path and the ideal-but-naive straight line.

The feasibility test on each bisection step is what makes the recipe concrete. For each new candidate path, the method determines "whether the new path meets an obstacle avoidance requirement and a structural constraint of the robot in the current scene." If it passes, the new (straighter) path "as the new planned path"; if it fails, the new path becomes "a new ideal path" — i.e., the algorithm backs off the target it is bisecting toward. It iterates "until an error between the planned path and the ideal path is within a preset range," then outputs the result as the robot's motion path. This is a specific shortcutting/path-smoothing procedure: start from a feasible sampled path, bisect toward the straight line, keep whatever stays collision-free and within the arm's structural limits, and converge. The "structural constraint" term is the tell that this is a manipulator method — it is reasoning about the arm's own geometry and reach, not just free-space obstacles.

“The present disclosure provides a robot motion path planning method, apparatus, and terminal device.”— U.S. Patent No. 10,821,605 source

The dependent claims push into joint space, which is where the manipulator framing becomes unmistakable. Claim 2 adds a second optimization stage that operates on "joint angle vector[s] of the joints of the robot": it picks the path point with "the greatest distance with respect to a corresponding point of the shortest ideal path," computes "an optimized joint angle vector" from "a preset scale factor, a shortest distance between the to-be-optimized point and the shortest ideal path, and a joint angle vector…at a point of the shortest ideal path," and accepts it only if it "meets the obstacle avoidance requirement and the structural constraint." Claim 4 even recites the formula — qopt2,imax = qideal,imax + w·dimax, where w < 1 is the scale factor and dimax the distance to the ideal path. A literal recited equation in a path-planning claim is unusual and unusually narrowing: the granted scope is tied to that joint-space update rule, not to optimization in the abstract. Claim 3 gives the stopping conditions (an "iteration number…larger than a preset iteration number threshold," or all points optimized).

For an actuation-beat analyst, the value of reading the classification and the claim together is that they tell you what prior art the claim must clear. A manipulator path-planning claim competes against decades of industrial-robotics literature — RRT, PRM, potential fields, and the standard "shortcut smoothing" post-processors that ship with OMPL itself. So the patentable delta is almost always in a specific computational step, and here it is explicit: the bisection-toward-the-ideal-path smoother plus the scale-factor joint-angle update of claims 2 and 4. The breadth lives in those recited steps; the title's generality is marketing, not scope. The supporting G05B 2219/404xx codes — safety, collision, and motion-control specifics in programmable controllers — confirm the claim is aimed at the execution layer, where a planned path becomes joint commands.

UBTECH is a consumer-and-humanoid robotics player, and a manipulator path-planning patent fits the portfolio of a company building arms and humanoids rather than self-driving systems. The terminal-device independent claim (claim 7) re-fences the identical OMPL-plus-dichotomy method as stored program instructions, so the enforceable surface spans method and device. Read alongside its other filings, this is portfolio-building in the B25J lane — the unglamorous control IP that any arm product needs and that competitors will eventually have to design around.

The caveat that always applies here is sharpened by the recited equation: a broad-sounding title ("motion path planning") almost never means a broad granted claim in a field this crowded, and a claim that names OMPL, a dichotomy method, and a specific joint-angle formula is narrow by construction. The B25J classification tells you the domain; claim 1's recited steps — bisection between the sampled and ideal paths under obstacle-and-structural constraints — tell you the fence. An accused system would have to use that bisection-smoothing scheme to read on it.

Bottom line for the manipulation beat: this is narrow, execution-layer arm IP — an OMPL-seeded, dichotomy-based path smoother with a scale-factor joint-space refinement — correctly classified and modestly scoped. Pull US10821605B2 in the patent record, read claim 1 with claims 2 and 4 for the bisection smoother and the joint-angle update, and track it as a data point in UBTECH's B25J velocity, not as a blocking patent on robot path planning writ large.