End-to-end robot intelligent obstacle avoidance method based on deep reinforcement learning with spatiotemporal...

Area of Science:

• Robotics
• Artificial Intelligence
• Computer Vision

Background:

• Current robot obstacle avoidance systems often have separate perception and decision-making modules, leading to fragmented data and poor generalization.
• Existing methods struggle with modeling dynamic environments and predicting obstacle trajectories effectively.

Purpose of the Study:

• To develop an end-to-end intelligent obstacle avoidance method for robots in complex dynamic environments.
• To improve autonomous decision-making, obstacle perception, and policy generalization capabilities.

Main Methods:

• Integration of Deep Q-Network (DQN) for reinforcement learning to learn optimal avoidance strategies.
• Incorporation of spatiotemporal attention mechanisms to model obstacle dynamics and collision risks.
• Design of a Transformer-based end-to-end architecture for integrated perception and decision-making using multi-modal inputs (LiDAR, depth images).

Main Results:

• The proposed method demonstrated superior performance in obstacle avoidance success rate, path optimization, and policy convergence speed compared to existing approaches.
• The system exhibited good stability and generalization capabilities in simulation environments.
• Eliminated the need for manual feature engineering and intermediate state modeling.

Conclusions:

• The integrated end-to-end approach offers a significant advancement in robot obstacle avoidance.
• The method shows strong potential for real-world applications in complex and dynamic environments.