SafeCrowdNav: safety evaluation of robot crowd navigation in complex scenes

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Mobile robot navigation in dense crowds is complex, requiring advanced collision avoidance.
• Current deep reinforcement learning methods lack quantitative safety metrics and can get stuck in local optima.
• Sparse rewards further limit learning optimal navigation strategies.

Purpose of the Study:

• To develop a comprehensive crowd navigation algorithm, SafeCrowdNav, for enhanced obstacle avoidance.
• To introduce a quantitative safety evaluation and an intrinsic exploration reward for improved learning.
• To address limitations of existing methods in complex crowd environments.

Main Methods:

• Proposed SafeCrowdNav algorithm integrating a safety evaluation function and an intrinsic exploration reward.
• Combined prioritized experience replay and hindsight experience replay for effective policy learning.
• Focused on improving robot comprehension of crowd dynamics.

Main Results:

• SafeCrowdNav demonstrated improved crowd comprehension during robot navigation.
• Significant reduction in collision probabilities compared to state-of-the-art algorithms.
• Shorter navigation times achieved with the proposed approach.

Conclusions:

• SafeCrowdNav offers a robust solution for safe and efficient robot navigation in crowded spaces.
• The algorithm's safety evaluation and exploration reward mechanisms are key to its success.
• Experimental results validate the effectiveness of SafeCrowdNav in real-world scenarios.