RAY-TOLD: Hybrid Control for Robot Navigation in Dense Crowds

A new hybrid control architecture, Ray-based Task-Oriented Latent Dynamics (RAY-TOLD), has been proposed to improve autonomous robot navigation through dense, dynamic crowds. The system integrates obstacle information into latent dynamics, combining physics-based Model Predictive Path Integral (MPPI) control with reinforcement learning for long-horizon foresight. RAY-TOLD uses a LiDAR-centric latent dynamics model to encode sensor data into a compact state representation, enabling a terminal value function and policy prior. A policy mixture sampling strategy augments MPPI candidate populations with learned policy trajectories, guiding the planner toward goals while avoiding local minima. The approach addresses persistent challenges in reactive planning methods like MPPI, which often fail in complex scenarios due to limited prediction horizons. The paper is available on arXiv under identifier 2604.27450.