Most research in deep learning focuses on a single modality, such as image, text, or proprioception data. However, humans benefit from leveraging information from diverse senses on a daily basis for richer information acquisition. Inspired by this, we design a transformer-based multimodal model for human action recognition and thoroughly evaluate its performance and robustness. Furthermore, we explore fusion methods to assess how modalities are best combined. Lastly, a model is trained to infer (generate) a missing modality. Our study shows that multimodal transformers perform better than their modality-specific equivalents. We achieve an improvement of \(10.1\%\) when using multiple data modalities over our vision-only baseline and outperform current state-of-the-art approaches by \(32.8\%\) . Furthermore, we evaluate a mean square error of \(9.6\%\) in the tactile force reconstruction task. The implemented model can be applied in scenarios where robotic assistance depends on recognising human actions for decision-making, tackling situations where vision is limited or audio and other modalities are required for deeper understanding.