GPTS-Nav: End-to-End Robot Navigation in Dynamic Environments With Graph-Privileged Teacher-Student Reinforcement Learning

Abstract

Reinforcement learning (RL) has renewed interest in end-to-end robot navigation, yet dynamic, crowd-like scenes remain difficult due to partial observability and the fragility of online tracking. We introduce GPTS-Nav, a Graph-Privileged Teacher-Student RL framework for LiDAR-only navigation. A graph-aware teacher is trained using single-step privileged motion graphs that explicitly include ground-truth velocities, forming a complete Markov state. In contrast, a sensor-only student relies purely on temporal observation history to implicitly infer unobservable dynamics. To enhance distillation fidelity and physical execution, a Gated LoRA Fusion module aligns student temporal features with teacher graph privileges, while an Action Chunking Transformer (ACT) predicts action chunks to filter jitter and surpass teacher trajectory smoothness. In large-scale simulations, GPTS-Nav consistently surpasses strong optimization-based and prior RL-based methods in success rate and efficiency. Furthermore, quantitative real-world experiments on a wheeled–biped platform validate its robust performance across narrow passages and uncooperative pedestrian zones, demonstrating safe, smooth navigation and readiness.