Load-Decomposition Learning Enhances Ground Reaction Force Estimation Using IMUs

Accurate estimation of ground reaction forces (GRFs) is fundamental to gait analysis and supports functional gait assessment, surgical planning, and rehabilitation. We propose a physics-learning hybrid framework that combines linked-segment dynamics with a bidirectional long short-term memory network to estimate vertical GRFs (vGRFs) from seven inertial measurement units (IMUs). The learning module produces an unconstrained allocation variable that decomposes the physics-derived net force into limb-specific vGRF…

By the Truvace deskOctober 15, 20261 min read1 sourcePeer-reviewedTRV-2026-1452· cite & verify

In brief

Researchers built a hybrid model that uses linked-segment dynamics to compute a net force and a bidirectional LSTM to split it into limb-specific vertical ground reaction forces. It takes input from seven inertial sensors worn on the body rather than force plates.

The result matters because accurate force estimates underpin gait assessment, surgical planning, and rehabilitation, including for people with transtibial amputation. The authors describe transfer across people, days, and tasks as potential, so real-world clinical benefit is still to be established.

Main points

The model pairs physics-derived net force with a learned allocation variable that splits force between the limbs.
Intra-subject RMSE was 0.034 BW with R2 of 0.994; inter-subject RMSE was 0.065 BW with R2 of 0.978.
It was compared against MLP, LSTM, BiLSTM, and Transformer baselines.
Data came from healthy participants, people with unilateral transtibial amputation, and a public biomechanics dataset.

The gain

Combining linked-segment physics with a BiLSTM network estimated vertical ground reaction forces from seven wearable IMUs with errors below a high-accuracy reference in intra- and inter-subject tests.

The rundown

The framework uses seven IMUs and a learning module that outputs an unconstrained allocation variable. That variable decomposes the physics-based net force into separate vertical force estimates for each limb.

Evaluation covered intra-subject, intersubject, cross-population, and between-day tests plus cross-task transfer. The public dataset included multiple walking speeds, ramps, and stairs.

The authors conclude that integrating biomechanical structure with data-driven learning can improve vGRF estimation and supports further development toward quantitative gait assessment and broader clinical applications.