Year: 2024
Type: conference-paper
Abstract: In the past few years, deep reinforcement learning has emerged as a novel way of manipulator path planning, so that the manipulator can realize the trajectory planning task in a high-dimensional complex scene, which makes up for the shortcomings of traditional motion planning in such scenarios, such as high computational complexity and difficulty in collision detection.However, many DRL models proposed for robots in complex scenes are difficult to obtain optimal strategies due to a large number of ineffective explorations and sparse rewards, which hinders the convergence of the model.Therefore, this paper proposes the following two methods: First, an inverse kinematics auxiliary strategy with guiding effect is proposed for the large number of invalid exploration problems in the DRL model.Second, the action strategy is adaptively selected by judging the distance from the obstacle to complete the obstacle avoidance.To assess the effectiveness of our proposed planning algorithm across various aspects, we carried out several trajectory planning experiments on the manipulator in intricate environments.The results demonstrate significant improvements in both the convergence speed and stability of trajectory planning using reinforcement learning in complex scenarios, owing to the two methods mentioned above.
Authors: Zhijian Meng, Xin Zhou, Hui Chen, Jianliang Mao
Institution: Shanghai University of Electric Power
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
Award: G1271446606
OpenAlex ID: W4411020767