Robotic Arm Trajectory Planning in Dynamic Environments Based on Self-Optimizing Replay Mechanism

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Robotic arms in dynamic environments face challenges like real-time changes and uncertainties.
• Deep reinforcement learning (DRL) for trajectory planning struggles with expert strategy acquisition, low experience utilization, and reward function design.

Purpose of the Study:

• To address the limitations of DRL in robotic arm trajectory planning.
• To improve convergence speed and performance in complex dynamic environments.

Main Methods:

• Designed a neural network-based expert-guided triple experience replay mechanism (NETM).
• Developed an improved reward function tailored for dynamic environments.
• Integrated imitation learning with DRL for optimized experience replay.

Main Results:

• NETM expands limited expert demonstrations and algorithm successes into optimized expert experiences.
• Experimental results demonstrate accelerated convergence in dynamic scenarios.
• NETM improved accuracy by over 30% and the safe rate by 2.28% compared to baseline algorithms.

Conclusions:

• The proposed NETM effectively enhances DRL for robotic arm trajectory planning.
• The approach significantly improves performance and convergence in dynamic environments.
• NETM offers a viable solution for real-world robotic applications facing uncertainty.