Human Activity Recognition (HAR) is a significant research area due to its many applications in various domains, including healthcare, sports, and human-computer interaction. In this paper, an approach to HAR using motion capture data to emulate wearable sensor data such as quaternions, angular velocities and accelerations is used. Long Short-Term Memory (LSTM) networks are used to classify the time series data representative of human movement patterns. To evaluate the model´s generalization capability, the evaluation was conducted in two stages: (1) training and testing the model with data collected under controlled conditions, and (2) assessing the model´s performance on data collected with variability in environmental conditions and calibrations, which creates a scenario closer to real-world conditions. Comparing stages, the results of the first stage demonstrate that the models achieve accuracy, precision, recall and F1 score close to 99%, indicating good performance in totally controlled conditions. However, the second stage performance degrades to around 88%, indicating the challenge of the classification model generalization in similar situations as real-word scenarios.