A new motion planner skips patchwork fixes and rebuilds broken robot routes wholesale.
Researchers built Masked Generative Motion Planning (MGMP), a system that generates multiple trajectory candidates in parallel using a masked generative transformer, then uses a technique called Geometry-Guided Token Search to decide what to edit and which alternatives to try. Instead of nudging a bad path with small local tweaks, the planner treats repair as a search problem over discrete route options, guided by the shapes and obstacles in the scene. In tests, it scored 96% success on a Ring Maze benchmark and 82% on a Controlled Route Invalidation task using a Kuka arm.
Those numbers matter because most robot planners still treat repair as an afterthought, nudging a bad path with gradient tweaks until something works. MGMP's results beat the strongest baselines by 23 and 25 percentage points, respectively, suggesting that rethinking repair as discrete search, rather than continuous smoothing, is a real improvement rather than a tuning trick. The system also generalized to new layouts, unseen obstacles and geometries, single- and dual-arm setups, and real-world tasks on a Baxter robot.
If local refinement has been the quiet bottleneck in generative motion planning, treating it as search instead of smoothing looks like the more durable fix. Whether it holds up outside lab mazes and arms on a bench is the next question.