Rice Roboticists Fold Perception Directly Into Motion Planning

A robot moving through a busy hospital corridor has to do two things at once that usually pull against each other: reach its goal, and keep the thing it cares about in view. Classical planners solve the first and treat the second as an afterthought, checking what the camera caught only after the path is drawn. A group of roboticists at Rice University argue that ordering is exactly backwards.

Their paper, Look as You Leap: Planning Simultaneous Motion and Perception for High-DOF Robots, was registered by IEEE Transactions on Robotics on 4 September 2026, with Crossref stamping the Early Access record at 19:14 UTC that day. The authors, Qingxi Meng, Emiliano Flores, Carlos Quintero-Peña, Peizhu Qian, Zachary Kingston, Shannan K. Hamlin, Vaibhav Unhelkar and Lydia Kavraki — have made the accepted version available on arXiv, where the comment field records it as accepted to the journal. The publisher's version of record sits behind IEEE's paywall; Unpaywall lists the DOI as closed, so the arXiv preprint is the readable primary for now.

Scoring perception inside the roadmap, not after it

The method the paper describes is a GPU-parallelized perception-score-guided probabilistic roadmap planner, which the authors call PS-PRM. According to the arXiv abstract, it uses a neural surrogate model to approximate perception scores, folds those scores into the roadmap search itself, and leans on GPU parallelism to replan online as a scene changes. The surrogate is what makes this affordable: scoring perception quality at every candidate node in a high-degree-of-freedom search space is expensive, and a learned approximation is fast enough to keep the planner running live.

That distinction, jointly optimizing motion and perception rather than sequencing them, is the whole bet. A path that reaches the goal but swings the camera away from a patient's face at the wrong moment is a failure the traditional pipeline only discovers after committing to it. Embedding the score in the search means the planner never proposes that path in the first place.

What the authors measured

In simulation on high-DoF Stretch robots navigating cluttered homes, the paper reports that its four PS-PRM variants reached an object detection rate above 70%, a confidence score exceeding 0.75, and a tracking rate around 90%, while classical baselines without perception in the planning loop often fell below 50% detection. The authors also report that the planner outperformed reinforcement-learning and trajectory-optimization baselines in both static and dynamic environments, across simulation and real-robot experiments. These are the authors' own experimental results from the accepted preprint, not an independent replication.

The framing is deliberately human-centered. The paper motivates the work with homes and hospitals, and its experimental sections include a nurse-training scenario using face detection on a Fetch robot, the kind of setting where continuous perception is not a nicety but the point of the deployment. A robot assisting in eldercare or clinical training that loses sight of the person it is tracking has not done a slightly worse job; it has stopped doing the job.

Why this is a planning problem, not a vision problem

What makes the result worth noting is where it locates the difficulty. Better cameras and better detectors do not help if the robot's own motion keeps pointing them at the wrong place. Perception quality, in this account, is partly a consequence of trajectory, and if it is, it belongs in the objective the planner optimizes rather than in a check that runs afterward. That reframing is more transferable than any single number in the table: it applies wherever a viewpoint has to be earned by moving well, which is most of the settings embodied robots are being pushed into.

The caveats are the ordinary ones for a fresh journal paper. The reported rates are the authors' measurements on their own hardware, Stretch in simulation, Fetch in the nurse-training face-detection scenario, and generality to other robots, other detectors and messier real deployments is not something a single paper can settle. Purdue's CoMMA Lab, home to co-author Zachary Kingston, still lists the work as "To Appear" even as the journal record exists. What the paper establishes is narrower and still useful: that treating perception as a first-class term in the planner, and paying for it with a learned surrogate cheap enough to run online, is a design that clears the baselines it was tested against.