Dual-Arm Robot Trajectory Planning Based on Deep Reinforcement Learning under Complex Environment.

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Dual-arm robots are increasingly used in healthcare settings.
• Complex environments, like those involving human patients and beds, pose significant collision risks.
• Sparse rewards in deep reinforcement learning can hinder robot task accomplishment.

Purpose of the Study:

• To develop an effective trajectory planning strategy for dual-arm robots in patient interaction scenarios.
• To address challenges of sparse rewards and collision avoidance in robot control.
• To train a neural network for precise manipulator control during patient handling tasks.

Main Methods:

• A 3D simulation environment was created to model realistic patient-robot interaction.
• A proximal policy optimization (PPO) algorithm with a continuous reward function was employed.
• A novel reward and punishment function, inspired by artificial potential fields, was designed, incorporating reward guidance, collision detection, obstacle avoidance, and time efficiency.

Main Results:

• The PPO algorithm demonstrated faster convergence, reducing training steps by approximately 4 million compared to the DDPG algorithm.
• The proposed reward and punishment function yielded significantly higher rewards within the same training time.
• The trained robot successfully reached target positions for patient handling, with shorter episode lengths and faster convergence.

Conclusions:

• The PPO algorithm, coupled with the novel reward function, offers a more efficient and effective approach to dual-arm robot trajectory planning for patient interaction.
• The developed method enhances safety by mitigating collision risks in complex environments.
• This research validates the advantages of the proposed algorithm for real-world robotic healthcare applications.