New POMDP Planner Boosts Robot Interception in Crowds by 31%

A team led by Himanshu Gupta at the University of Colorado Boulder has introduced a new online POMDP (Partially Observable Markov Decision Process) planning method for robots intercepting moving targets in crowded human environments. The robot must reason over multiple possible human navigation intents, which are not directly observable. The researchers formulated this as a POMDP solved with online tree search under a fixed computational budget. They compared two action-space structures: a sequential path-speed planner that first plans a spatial path then modulates speed, and a unified planner that jointly branches over steering and speed within the tree search.

In simulations with up to 200 virtual humans, both approaches performed similarly at low density but diverged sharply as density increased. At the highest crowd density, the sequential planner achieved a safe-interception rate 31 percentage points lower than the unified planner and required 44% more time to complete the task. This reveals a structural limitation of spatial restriction in dense crowds. The work, accepted at IROS 2026, demonstrates that allowing simultaneous reasoning over both path and speed decisions is critical for reliable robot navigation in crowded spaces, with direct applications to delivery robots, service robots, and autonomous systems operating in busy public areas.