Diversity-Aware Crowd Model for Robust Robot Navigation in Human Populated Environment
Jiaxu Wu, Yusheng Wang, Tong Chen, Jun Jiang, Yongdong Wang, Qi An, Atsushi Yamashita
AI summary
Problem
Existing RL-based robot navigation methods rely on crowd simulators that lack behavioral diversity, causing trained policies to overfit to specific human patterns and fail in unseen real-world scenarios.
Approach
The authors propose a diversity-aware crowd model trained via Reinforcement Learning that uses Constrained Variational Exploration with a Mutual Information-based auxiliary reward and a Centralized Training Decentralized Execution (CTDE) paradigm to generate fine-grained, near-optimal human navigation behaviors without pre-collected data.
Key results
- Novel RL framework integrating Constrained Variational Exploration and CTDE for diverse crowd simulation
- Collection and evaluation of a new real-world human navigation dataset
- Superior robot navigation robustness in simulation and real-world tests compared to rule-based and RL baselines
- Stabilized training and higher success rates in crowded scenarios through optimality constraints and centralized critics
Why it matters
Enables safer, more reliable autonomous robot deployment in complex human environments by overcoming the sim-to-real diversity gap.
Abstract
Robot navigation in human-populated environments poses challenges due to the diversity of human behaviors and the unpredictability of human paths. However, existing Rein- forcement Learning (RL)-based methods often rely on simulators that lack sufficient diversity in human behavior, resulting in navigation policies that overfit specific human behavior and perform poorly in unseen environments. To address this, we propose a diversity-aware crowd model based on RL, employ- ing Constrained Variational Exploration (VE) with a Mutual Information (MI)-based auxiliary reward to capture fine-grained behavioral diversity. The proposed model leverages a Centralized Training Decentralized Execution (CTDE) paradigm, which en- sures stable exploration under multi-agent settings. Using the pro- posed diversity-aware model for training, we obtain robust robot navigation policies capable of handling diverse unseen scenarios. Extensive simulation and real-world experiments demonstrate the superior performance of our approach in achieving diverse crowd behaviors and enhancing robot navigation robustness. These findings highlight the potential of our method to advance safe and efficient robot operations in complex dynamic envi- ronments. For more details, please visit our project homepage https://wyd0817.github.io/project-diversity-awa/.