Motion planning framework based on dual-agent DDPG method for dual-arm robots guided by human joint angle...

Area of Science:

• Robotics
• Artificial Intelligence
• Motion Planning

Background:

• Reinforcement learning (RL) is prevalent in robot motion planning.
• Existing RL methods face challenges in complex multi-step tasks for dual-arm robots, including large exploration spaces, extended training durations, and limited process control.
• These limitations hinder efficient and human-like robot behavior.

Purpose of the Study:

• To develop an improved motion planning framework for dual-arm robots.
• To address the limitations of current RL-based trajectory planning methods.
• To achieve humanized learning content and style while enabling rapid, coordinated trajectory planning for complex multi-step tasks.

Main Methods:

• The framework utilizes the dual-agent deep deterministic strategy gradient (DADDPG) algorithm.
• It incorporates human joint angle constraints derived from human arm motion data (IMU) and a human-robot kinematic mapping model.
• A segmented reward function, guided by these constraints, is designed to reduce exploration space and accelerate training.

Main Results:

• The framework was validated using the Baxter robot in a reach-grasp-align task simulation.
• Results demonstrate that human experience knowledge significantly influences learning guidance.
• The proposed method enables faster planning of coordinated dual-arm trajectories for multi-step tasks.

Conclusions:

• The developed motion planning framework effectively integrates human joint angle constraints into RL.
• This approach enhances the efficiency and controllability of dual-arm robot motion planning for complex tasks.
• The findings highlight the potential of human-guided RL for more intuitive and rapid robot learning.