Estimating lower-limb joint torques during gait using inertial measurement units (IMUs) has attracted growing attention in biomechanics and wearable sensing. Conventional approaches rely on inverse dynamics based on segmental kinematics and ground reaction forces, requiring force sensors or full-body sensor setups. This study proposes a recurrent neural network (RNN) method to estimate lower-limb joint torques using segmental kinematic data from a limited number of IMUs.Twelve healthy participants performed treadmill walking and running under twelve different conditions to generate training data. Model inputs included center-of-mass accelerations and angular velocities of the pelvis and shank.Results demonstrated two key findings. First, a model using three IMUs achieved performance comparable to a seven-IMU model, with hip flexion torque errors of approximately 0.18 Nm/kg, demonstrating strong effectiveness with a reduced sensor configuration. Second, while inverse dynamics exhibited an error increase of 0.28 Nm/kg from the ankle to the hip, the proposed model showed only a 0.01 Nm/kg increase and achieved approximately 0.13 Nm/kg lower error at the hip.These results indicate that accurate and efficient joint torque estimation is feasible using an RNN with fewer wearable sensors.
