Open Access

ARTICLE

LSTM-Enhanced Deep Reinforcement Learning for Active Motion Compensation of Surgical Robots with Known Target Position

Wei Wei^1,2, Shujuan Li¹, Legend Zhang³, Junmin Lyu³, Qi Hu⁴, Wenfeng Zheng^3,4,*, Bo Yang^4,*

1 School of Optoelectronic Science and Intelligent Instrumentation, Xi’an University of Technology, Xi’an, China
2 School of Intelligent Manufacturing, Xi’an University, Xi’an, China
3 Future Tech Institute, Guangzhou Huashang University, Guangzhou, China
4 School of Automation Engineering, University of Electronic Science and Technology of China, Chengdu, China

* Corresponding Authors: Wenfeng Zheng. Email: ; Bo Yang. Email:

(This article belongs to the Special Issue: Recent Advances in Signal Processing and Computer Vision, 2nd Edition)

Computer Modeling in Engineering & Sciences 2026, 148(2), 45 https://doi.org/10.32604/cmes.2026.085095

Received 05 May 2026; Accepted 29 July 2026; Issue published 28 August 2026

Abstract

Active motion compensation is essential for improving the precision and safety of robot-assisted surgery in the presence of physiological motion such as heartbeat and respiration. Conventional direct error feedback controllers often show limited performance when sensing delay and measurement noise are present. To address this issue, this study proposes an active motion compensation framework based on deep reinforcement learning enhanced with a Long Short-Term Memory (LSTM) network, where the target position is assumed to be known. The motion compensation task is formulated as a Markov decision process, and the controller is trained to generate continuous control forces for the surgical instrument in three Cartesian directions. To overcome the performance degradation caused by sensing delay, LSTM is incorporated into both the actor and critic networks. Four controllers, including DDPG, TD3, DDPG-LSTM, and TD3-LSTM, are trained and evaluated in a Unity-based simulation environment with a 40 ms sensing delay and zero-mean Gaussian measurement noise. Experiments are conducted on simulated motion signals as well as Phantom and in vivo cardiac motion datasets. The results show that conventional fully connected controllers trained by DDPG and TD3 fail to achieve satisfactory compensation under delayed conditions, whereas the proposed LSTM-enhanced controllers significantly improve tracking performance. Among them, TD3-LSTM achieves the best overall results, with RMSE values of 0.1739, 0.3823, and 0.6136 mm on simulated, Phantom, and in vivo data, respectively. In addition, the proposed LSTM-enhanced controller outperforms a conventional PD controller, demonstrating the effectiveness of combining temporal sequence modeling with reinforcement learning for delay-aware surgical motion compensation.

---

Keywords

---