The dynamic analysis of lower limb biomechanics is crucial for understanding gait, posture, and load distribution, which are foundational for controlling assistive robots like exoskeletons and intelligent prostheses. Traditional methods, including invasive musculoskeletal measurements, while providing precise data, are costly, intrusive, and technically complex, limiting their widespread application. To overcome these limitations, noninvasive approaches, such as musculoskeletal multibody dynamics simulations (MMDS), have been proposed. These simulations combine data from noninvasive sensors like motion capture systems and force plates to model the internal forces and moments of the body. “However, MMDS frameworks face issues related to their dependence on force plates, their high computational demand, and their limited ability to provide real-time feedback, which is essential for dynamic applications such as robot control.” said the author Hao Zhou, a researcher at Shenzhen Institutes of Advanced Technology, “Therefore, we propose a lightweight and computationally efficient deep learning model, the Marker GMformer, to address these challenges. This model integrates prior anatomical knowledge and spatiotemporal features, and can continuously predict lower limb biomechanical data in real-time.”
Marker-GMformer is a lightweight deep learning model specifically designed to predict multi-joint kinematics, joint torques, and three-dimensional ground reaction forces (GRFs) of the lower limbs from labeled trajectory data. The architecture of Marker-GMformer primarily consists of a temporal block, a spatial block, and a temporal embedding layer. The spatial block combines a Graph Convolutional Network (GCN) and a Multi-Layer Perceptron (MLP) module to extract local and global features of the lower limb anatomy. By constructing an adjacency matrix that reflects skeletal connectivity, GCN can capture spatial correlations between markers and integrate anatomical prior information. The temporal block employs an improved Transformer model to extract global temporal features from the marker coordinate time series. Through the ProbSparse self-attention mechanism, redundant computations are reduced, thereby enhancing computational efficiency and enabling the processing of longer time series data. In traditional Transformer models, time series are typically encoded by time steps, whereas in Marker-GMformer, each marker of the entire time series is treated as a separate token and embedded through a fully connected layer. This approach enhances the model's ability to extract temporal features. Simultaneously, by facilitating information flow between the spatial and temporal blocks, Marker-GMformer can effectively learn spatial-temporal dependencies, accurately predicting kinematics, dynamic variables, and ground reaction forces of the lower limbs.
The Marker-GMformer model successfully predicted lower limb multi-joint angles, joint moments, and 3D ground reaction forces (GRFs) across 13 different motion patterns. The predicted results were highly consistent with those from traditional musculoskeletal dynamics simulations (MMDS) and force plate measurements, showing excellent accuracy: The correlation coefficients for all predicted variables were greater than 0.97, indicating a strong linear relationship between the predictions and actual data. The RMSE for joint angles was 1.95°, joint moments was 0.099 N·m/kg, and for GRFs was 0.036 body weight, indicating the model's high prediction accuracy. The model had low computational complexity, enabling real-time inference suitable for applications requiring fast feedback (such as robotic control and real-time lower limb biomechanics monitoring). Marker-GMformer performed exceptionally well in various motion patterns, particularly in walking, running, and inclined walking, where the prediction accuracy was highest. However, for more dynamic movements like squatting, vertical jumping, and hopping, the accuracy of the predictions was lower, especially for some moments (like hip moments) and ground reaction forces (APGRF), where larger errors were observed.
Overall, Marker-GMformer model has significant advantages, especially in its feasibility for real-time applications. Compared to traditional musculoskeletal dynamics simulations (MMDS), Marker-GMformer only relies on marker trajectory data, eliminating the need for force plates or complex modeling platforms, greatly simplifying data collection and computation. By combining anatomical prior knowledge with spatiotemporal feature extraction, Marker-GMformer performs excellently across various motion patterns, providing high-accuracy predictions while reducing computational load, making it ideal for real-time feedback and control tasks, such as robotic control and lower limb biomechanics monitoring. “To further enhance the accuracy of Marker-GMformer, particularly its performance in dynamic tasks, we will explore enhancing the diversity of the model's dataset in the future by incorporating more data on extreme dynamic or non-periodic movements. Additionally, by incorporating physical constraints or biomechanical prior knowledge to ensure a smooth transition of torques and ground reaction forces, the physical rationality and stability of the model may be further improved.” said Hao Zhou.
Authors of the paper include Hao Zhou, Yinghu Peng, Xiaohui Li, Xueyan Lyu, Hongfei Zou, Xu Yong, Dahua Shou, Guanglin Li, and Lin Wang.
This work was supported by the National Key Research and Development Program of China [grant number 2024YFE0216500]; the Shenzhen Strategic Emerging Industry Support Plans [grant number XMHT20230115002]; the Shenzhen Sustainable Development Sci-Tech project [grant number KCXFZ20230731093501003]; the Shenzhen Science and Technology Program [grant number KQTD20210811090217009]; and the Shenzhen Science and Technology Program [grant number JCYJ20240813154923031].
The paper, “Continuous Lower Limb Biomechanics Prediction via Prior-Informed Lightweight Marker-GMformer” was published in the journal Cyborg and Bionic Systems on Jan 15, 2026, at DOI: 10.34133/cbsystems.0476.
END