By developing an add-on that can sit on top of most electric wheelchair joysticks instead of working with direct wiring of a specific wheelchair model, we aim to make our solution more accessible and customizable.

The initial prototype uses three classes: feet (forward), right hand (rotate right), and rest. This is a deliberate choice to minimize system complexity and prioritize a functional, usable prototype over premature feature expansion. The final product will support will allow the user to either choose which hand should be used to rotate the wheelchair, or to use both hands to rotate on both sides: feet (forward), right hand (rotate right), left hand (rotate left), and rest.

[add short description about the joystick cap]

We begin with conventional classification models such as LDA and SVM with CSP features, as they are lightweight and predictable. If these hit a wall in accuracy, we will move to deep learning approaches like CNNs or transformers.

All models are benchmarked against MOABB (Mother of All BCI Benchmarks) for standardized comparison.

[add description here]

Next steps: - Allow for continuous input alongside discrete commands (for example, one command to go forward by 8 feets instead of 4 commands of moving 2 feets forward)