Figure 2. The general architecture of the GRF estimation models.

The model is made up of the two-dimensional time series input matrix for each foot contact, being passed through the bi-directional LSTM layer, after which the hidden state vectors from each time point are passed through three linear layers to map the hidden state vectors to a single predicted force value. The size of the layers and activation functions being used were identified through hyperparameter optimisation, results shown in Table 2.