Offline-Trained GAN-Augmented Highly Adaptive Control with Multi-DoF Fusion for Pneumatic Soft Surgical Robots

Pneumatic soft robots are well-suited for minimally invasive surgery owing to their compliance and safe interac- tion with tissues. However, achieving highly adaptive control is difficult owing to modeling inaccuracies, inter-chamber cou- pling, and disturbances from surgical instruments. Non-learning adaptive methods depend on simplified models and perform poorly in unstructured settings. Conversely, learning-based methods often impose high computational costs in multi- degree-of-freedom (multi-DoF) pneumatic systems. A previous study proposed a generative adversarial network (GAN)-based proportional–integral–derivative (G-PID) controller that com- bined PID stability with learning-based adaptability by aligning system behavior with a reference model. However, its perfor- mance in highly coupled multi-DoF pneumatic soft robots was unverified, and its online adversarial training was computa- tionally intensive. We addressed these limitations by developing an offline-trained G-PID controller, shifting adversarial training offline to reduce computational overhead, achieving 23-fold faster convergence, and enabling real-time, model-free control with balanced adaptability and efficiency. We evaluated three multi- DoF data fusion strategies, showing effective coordination of DoF coupling while maintaining individual control fidelity. Validation on a multi-DoF soft robotic mechatronic system for single-port transvesical prostatectomy revealed tip errors below 0.16 mm across surgical instruments. Proposed controller enhances scala- Received 19 September 2025; revised 10 November 2025; accepted 7 December 2025. Date of publication 16 December 2025; date of current version 12 January 2026. This article was recommended for publication by Associate Editor H. Wang and Editor L. Zhang upon evaluation of the reviewers’ comments. This work was supported by the National Natural Science Foundation of China under Grant 62088101. (Corresponding author: Zhongchao Zhou.) Yuxi Lu is with Shanghai Research Institute for Intelligent Autonomous Systems, the State Key Laboratory of Autonomous Intelligent Unmanned Systems, and the Frontiers Science Center for Intelligent Autonomous Systems of the Ministry of Education, Tongji University, Shanghai 201203, China, also with Shanghai Artificial Intelligence Laboratory, Shanghai 200232, China, and also with the Center for Frontier Medical Engineering, Chiba University, Chiba 263-8522, Japan (e-mail: yuxilu@tongji.edu.cn). Zhongchao Zhou is with the School of Engineering, The University of Tokyo, Tokyo 113-0033, Japan (e-mail: zhouzhongchao@outlook.com). Dongliang Zheng, Yanmin Zhou, Zhipeng Wang, Shuo Jiang, and Bin He are with Shanghai Research Institute for Intelligent Autonomous Systems, the State Key Laboratory of Autonomous Intelligent Unmanned Systems, and the Frontiers Science Center for Intelligent Autonomous Systems of the Ministry of Education, Tongji University, Shanghai 201203, China, and also with Shanghai Artificial Intelligence Laboratory, Shanghai 200232, China (e-mail: dzheng46@tongji.edu.cn; yanmin.zhou@tongji.edu.cn; wangzhipeng@tongji.edu.cn; jiangshuo@tongji.edu.cn; hebin@tongji.edu.cn). Wenwei Yu is with the Center for Frontier Medical Engineering, Chiba University, Chiba 263-8522, Japan (e-mail: yuwill@faculty.chiba-u.jp). This article has supplementary downloadable material available at https://doi.org/10.1109/TASE.2025.3644871, provided by the authors. Digital Object Identifier 10.1109/TASE.2025.3644871 bility and adaptability and may generalize to other mechatronic systems with nonlinear, coupled dynamics. Note to Practitioners—Pneumatic soft manipulators are attrac- tive for minimally invasive procedures; however, they are difficult to control in practice: air chambers interact, simple models are inaccurate, and controllers that learn online can be too slow for real-time use. This work leverages a familiar PID loop and adds a compact compensator trained offline on short motion trials. In operation, the controller behaves like standard real-time system —no online learning or detailed physical model—while correcting for coupling and tool-induced disturbances. We also co mpared three multi-DoF fusion strategies and found that a unified fusion of all DoFs worked best. For a complete multi-DoF soft robotic mechatronic system designed for a single-port prostatectomy task, the approach maintained a tip error of less than 0.16 mm across different instruments, thereby reducing retuning effort and improving repeatability. Deployment involves collecting brief device-specific trajectories, training the compensator offline, and integrating it with existing pressure regulation and motion track- ing. Current limitations include testing mainly planar motions on a four-chamber device, response speed constrained by pneumatic hardware, and the need for formal stability guarantees and automated tuning; higher-flow valves, revised chamber geometry, compact networks, and stability certification are practical next steps. Beyond surgery, the method is expected to transfer to other nonlinear, coupled mechatronic systems (e.g., continuum robots, pneumatic-muscle actuators, and compliant end-effectors).