Social Robot Navigation Using Subgoal-Based MPPI for Crowd Flow Guidance

다국어 초록 (Multilingual Abstract)

This thesis proposes a hierarchical framework that enables a mobile robot to navigate in a socially acceptable manner in dense pedestrian environments by explicitly accounting for crowd flow. Conventional approaches rely primarily on the positions and velocities of individual pedestrians and therefore do not leverage the macroscopic crowd flow collectively formed by multiple pedestrians. As a consequence, the robot tends to move into flows that oppose its goal direction, resulting in repeated avoidance and stopping behaviors—a phenomenon referred to as the counter crowd flow problem. The proposed framework consists of a high-level subgoal selector based on reinforcement learning and a low-level Model Predictive Path Integral （MPPI） tracker. The high-level policy evaluates candidate subgoals around the robot using goal progress, the direction and magnitude of the local crowd flow, crowd density, and the minimum distance to obstacles and pedestrians. By incorporating the alignment between the crowd flow and the goal direction into the reward function, the policy selects an appropriate subgoal, which the MPPI tracker subsequently tracks as a short-term target while handling collision avoidance. Experiments are conducted across three scenarios and three crowd sizes （10, 20, and 30 pedestrians）, with each of the nine conditions evaluated over 500 test cases. Across all nine conditions, the proposed method attains a success rate of 0.92–1.00 and a collision rate of 0.00–0.06, while keeping the intrusion time ratio within 0.02–0.64 %. It runs in real time with an inference time of approximately 1 ms per step, on the order of 20 times faster than the most computationally intensive baseline. Finally, an ablation analysis based on the Counterflow Crossing Index （CCI）, a metric introduced to quantify how far the robot travels against the surrounding crowd flow, confirms that the observed performance gains result from incorporating crowd flow into the subgoal selection process rather than from the subgoal structure itself

목차 (Table of Contents)

• 1 Introduction 1
• 2 Related Work 6
• 2.1 Reinforcement Learning-Based Crowd Navigation 6
• 2.2 Hierarchical Navigation and Subgoal-Based Approaches 8
• 2.3 Crowd-Flow-Aware Navigation 10
• 3 Proposed Method 11
• 3.1 Problem Formulation 13
• 3.2 Candidate Subgoal Generation and Feature Representation 17
• 3.3 Crowd-Flow-Aware Subgoal Policy Network 23
• 3.4 Reward Design 27
• 3.5 MPPI-Based Low-Level Controller 30
• 4 Experimental Design 35
• 4.1 Simulation Environment 35
• 4.2 Experimental Scenarios 36
• 4.3 Implementation Details 37
• 4.4 Baseline Algorithms 39
• 4.5 Evaluation Metrics 39
• 5 Experimental Results 42
• 5.1 Quantitative Comparison 42
• 5.2 Metric Analysis 43
• 5.3 Scenario Analysis 45
• 5.4 Qualitative Analysis 48
• 5.5 Computational Efficiency 50
• 5.6 Ablation Study 52
• 5.7 Discussion on Real-World Applicability 55
• 6 Conclusion 57
• References 59