Sampling-Based Motion Planning with Scene Graphs Under Perception Constraints

A research team led by Qingxi Meng from Rice University, with collaborators from other institutions, has introduced MOPS-PRM, a novel motion planning algorithm that addresses a critical gap in robotics. While robots increasingly operate in cluttered human environments like homes and hospitals, they must maintain perception constraints—keeping people or objects in view for safety—while executing tasks. Existing methods either handle low-degree-of-freedom systems or focus on single objects, leaving high-DoF robots (like manipulator arms) struggling with multi-object monitoring. MOPS-PRM tackles this by integrating perception costs directly into a scene graph representation of the environment, which stores rich semantic and spatial data but previously lacked perception-aware information.

The technical innovation lies in embedding a 'perception cost' for observing each object into the scene graph. This cost guides the Probabilistic Roadmap (PRM) sampling process, selectively generating robot configurations that favor good viewpoints. The method was validated in simulation and real-world experiments, showing a 36% improvement in the average number of detected objects and a 17% better track rate compared to other perception-constrained baselines, all while maintaining similar planning times and path lengths. This work, accepted to IEEE Robotics and Automation Letters (R-AL), bridges task planning and motion planning, enabling robots to perform long-horizon tasks safely by continuously monitoring their surroundings. The next steps involve scaling to more dynamic environments and integrating with advanced vision models.