A deep reinforcement learning framework for crowd-aware robot navigation using attention mechanisms to model human-robot and human-human interactions.
CrowdNav is a deep reinforcement learning framework for crowd-aware robot navigation that models both human-robot and human-human interactions. It enables robots to navigate efficiently and socially-compliantly in dense crowds by using attention mechanisms to anticipate human dynamics. The framework addresses the limitation of existing methods whose cooperation ability deteriorates as crowd density increases.
Robotics researchers and developers working on autonomous navigation in human-populated environments, particularly those interested in social robot navigation and deep reinforcement learning applications.
It provides a more comprehensive approach to crowd navigation by explicitly modeling crowd-robot interactions rather than just pairwise human-robot interactions, leading to better anticipation of human dynamics and more efficient navigation in dense crowds compared to state-of-the-art methods.
[ICRA] Crowd-aware Robot Navigation
Open-Awesome is built by the community, for the community. Submit a project, suggest an awesome list, or help improve the catalog on GitHub.
Uses self-attention to model both Human-Robot and Human-Human interactions, improving anticipation in dense crowds as validated in the ICRA 2019 paper.
Outperforms state-of-the-art methods like ORCA and SARL in time efficiency and success rates, demonstrated through simulation videos and learning curves in the README.
Includes a customizable gym_crowd environment for training and testing, facilitating research reproducibility and benchmarking with detailed setup instructions.
Fully open-source with paper citations, code availability, and comparison to baselines, ensuring academic credibility and ease of extension.
Primarily optimized for simulation environments; integrating with real robots requires significant additional engineering for sensor fusion and real-time processing.
Based on 2019 research, with newer works like Relational Graph Learning referenced, indicating potential obsolescence for cutting-edge applications without updates.
Requires installation of niche libraries like Python-RVO2 and custom pip installs, which can be error-prone and time-consuming for users unfamiliar with the ecosystem.