Ground reaction force (GRF) is essential for estimating muscle strength and joint torque in inverse dynamic analysis. Typically, it is measured using a force plate. However, force plates have spatial limitations, and studies of gaits involve numerous steps and thus require a large number of force plates, which is disadvantageous. To overcome these challenges, we developed a deep learning model for estimating three-axis GRF utilizing shoes with three uniaxial load cells. GRF data were collected from 81 people as they walked on two force plates while wearing shoes with three load cells. The three-axis GRF was calculated using a seq2seq approach based on long short-term memory (LSTM). To conduct the learning, validation, and testing, random selection was performed based on the subjects. The 60 selected participants were divided as follows: 37 were in the training set, 12 were in the validation set, and 11 were in the test set. The estimated GRF matched the force plate-measured GRF with correlation coefficients of 0.97, 0.96, and 0.90 and root mean square errors of 65.12 N, 15.50 N, and 9.83 N for the vertical, anterior–posterior, and medial–lateral directions, respectively, and there was a mid-stance timing error of 5.61% in the test dataset. A Bland–Altman analysis showed good agreement for the maximum vertical GRF. The proposed shoe with three uniaxial load cells and seq2seq LSTM can be utilized for estimating the 3D GRF in an outdoor environment with level ground and/or for gait research in which the subject takes several steps at their preferred walking speed, and hence can supply crucial data for a basic inverse dynamic analysis.

Traditional force plate-based systems offer high measurement precision but are limited to laboratory settings, restricting their use in real-world environments. To address this, we propose a method for estimating a three-axis ground reaction force (GRF) and two-axis center of pressure (CoP) using a shoe embedded with three uniaxial load cells. The estimation was conducted under five gait conditions: straight walking, turning, uphill, downhill, and running. Data were collected from 40 healthy young adults. Four deep-learning models—Fully Connected Neural Network (FCNN), Convolutional Neural Network (CNN), Sequence-to-Sequence Long Short-Term Memory (Seq2Seq-LSTM), and Transformer—were evaluated. Among them, Seq2Seq-LSTM and CNN achieved the highest performance in predicting both GRF and CoP. However, the medio-lateral (ML) components showed lower accuracy than the vertical and anterior–posterior directions. In slope conditions, particularly for vertical GRF, relatively higher root mean-square error (RMSE) values were observed. Despite some variation across gait types, predicted values showed high agreement with measurements. Compared with previous studies, the proposed method achieved comparable or better performance with a minimal sensor setup. These findings highlight the feasibility of accurate GRF and CoP estimation in diverse gait scenarios and support the potential for real-world applications. Future work will focus on sensor optimization and broader population validation.