A Momentum Recurrent Neural Network for Sparse Motion Planning of Redundant Manipulators With Majorization-Minimization
Authors: Haoen Huang, Long Jin, Zhigang Zeng
Abstract:
Redundant manipulators possess more degrees of freedom than are strictly required to accomplish an end-effector task. This redundancy gives them flexibility for avoiding obstacles, respecting physical constraints, and optimizing secondary objectives. However, it also makes real-time motion planning computationally demanding. Two practical issues remain especially important: conventional recurrent-neural-network solvers may converge too slowly for time-critical robotic applications, and most existing approaches do not explicitly promote sparse joint motion. Sparse motion planning is desirable because accomplishing a task with fewer actively moving joints can simplify coordination, reduce unnecessary actuation, and improve the efficiency and interpretability of the resulting motion.
To address these issues, this paper proposes a momentum recurrent neural network (MRNN) inspired by Nesterov’s accelerated gradient method. By introducing a momentum mechanism into the neural dynamics, the MRNN exploits both the current optimization direction and accumulated dynamic information, thereby accelerating convergence while retaining the online-computing advantages of recurrent neural networks. The resulting model provides an efficient neurodynamic solver for the time-varying optimization problems that arise in redundant-manipulator motion planning.
Sparse motion planning introduces a nonconvex objective because it seeks to minimize the number of active joints rather than merely the magnitude of joint velocities. The paper handles this challenge through a majorization-minimization strategy. By this way, the sparse optimization problem is transformed into a nonlinear system. Multiple MRNNs then operate concurrently within a collaborative neurodynamic optimization (CNO) framework. Through parallel exploration and collaboration among the neural solvers, the proposed CNO-MRNN method searches effectively for a high-quality sparse solution while satisfying the manipulator’s tracking task and relevant constraints.
The method is evaluated through both computer simulations and physical robot experiments. The results show that the proposed MRNN improves time efficiency and trajectory-tracking performance compared with competing neural-network approaches. In the reported sparse motion-planning tests, CNO-MRNN achieves the best overall tracking performance among the compared methods, with a maximum position error of 2.34×10-3 m and a mean-square error of 1.51×10-3 m, while using only four active joints. These results demonstrate that accelerated neural dynamics and sparsity-oriented optimization can be combined successfully for accurate and efficient motion planning of redundant manipulators.
The work offers a general perspective on intelligent robotic optimization: momentum can accelerate continuous-time neural solvers, while majorization-minimization and collaborative neurodynamic optimization can make difficult nonconvex sparse-planning problems practically solvable. The proposed framework therefore has potential value for robotic systems in which fast online response, accurate end-effector tracking, and economical joint coordination are simultaneously required, including industrial manipulation, service robotics, and operation in constrained environments.
Additional Information
This work demonstrates that a momentum-accelerated recurrent neural network, combined with majorization-minimization and collaborative neurodynamic optimization, can deliver accurate trajectory tracking while activating only a sparse subset of a redundant manipulator’s joints.
- This paper introduces a momentum recurrent neural network (MRNN) inspired by Nesterov’s accelerated gradient method to improve the convergence speed.
- This paper develops a collaborative neurodynamic optimization framework in which multiple MRNNs work concurrently to solve a nonconvex sparse motion-planning problem.
- This paper validates the method through both computer simulations and physical robot experiments.
- This paper connects accelerated optimization, recurrent neural dynamics, and sparse robotic motion planning in a unified framework.
The research team of Haoen Huang, Long Jin, and Zhigang Zeng works at the intersection of neural dynamics, optimization, and intelligent robotic control. Their research seeks computationally efficient neural approaches for real-time optimization and motion control, with particular emphasis on redundant manipulators and constrained robotic systems.
Haoen Huang
- Position: Postdoctoral Researcher
- Affiliation: School of Artificial Intelligence and Automation, Huazhong University of Science and Technology, Wuhan 430074, China
- Email: henhuang@hust.edu.cn
Long Jin – corresponding author
- Position: Professor
- Affiliation: School of Artificial Intelligence, Lanzhou University, Lanzhou 730000, China
- Public contact email: jinlong@lzu.edu.cn
Zhigang Zeng – corresponding author
- Position: Professor
- Affiliation: School of Artificial Intelligence and Automation, Huazhong University of Science and Technology, Wuhan 430074, China
- Public contact email: zgzeng@hust.edu.cn
Laboratories and research groups
- School of Artificial Intelligence and Automation, Huazhong University of Science and Technology. Website: https://aia.hust.edu.cn/
- School of Artificial Intelligence, Lanzhou University. Website: https://ai.lzu.edu.cn/
- Brain-Inspired Intelligence and Intelligent Control (BIIC) Research Group, led by Prof. Zhigang Zeng. Website: http://biic.aia.hust.edu.cn/
- Joint Intelligence and Network Laboratory, led by Prof. Long Jin. Website: https://jinl.lzu.edu.cn/
Funding acknowledgment
The work was supported in part by the National Key R&D Program of China under Grant 2021ZD0201300, the 111 Project on Computational Intelligence and Intelligent Control under Grant B18024, and the Foundation for Outstanding Research Groups of Hubei Province of China under Grant 2025AFA012.
Demonstration video and other resources
- Paper DOI: https://doi.org/10.1109/TIE.2025.3566731
- Demonstration video: https://youtu.be/pEsK4E3aV4w