Beyond the Joystick

Update 3 · May 23rd 2026

The chair moved.

In This Update

Hardware: actual chair on the floor, not pixels on a screen. The full pipeline from EEG signal to physical motor command is now operational.

What Happened Today

We started the day with the hardware integration complete and the safety hardware in place. The PCAN-USB interface was wired to R-Net Port 3 on the chair, the fault-tolerant transceiver was in line, the independent hardware emergency stop was installed and tested, and the chair-side software was a blank file.

By the end of the day, the chair was driving from intent.

The sequence:

We verified the chair's live bus traffic matched our captured protocol. The chair produces master drive frames at 100 Hz and the JSM produces joystick frames at 100 Hz across a range of -100 to +100 on each axis. Our captures match the live chair exactly. We then verified the inject path by sending harmless test frames on an unused arbitration ID and confirming the chair's bus echoed them back to our listener. Read and write paths proven independently.

Forward motion at conservative magnitude on a chair lifted off the ground produced clean rotation of both drive wheels for sustained periods. We then ran a series of direction tests to map our outputs to the chair's actual response, captured ground-truth byte layout directly from the original joystick module, and corrected our injection format.

Then we hooked up the Consumer EEG

The first BCI-to-chair session ran for 67 seconds. The model fired five forward commands and four right-turn commands. Each one produced physical chair motion. Total time from putting on the headband, starting the script, and driving: under one minute.

What This Demonstrates

This is the moment we have been working towards for over a year. The simulator showed that the upstream half of the system was sound. Today showed that the downstream half works on the actual hardware the project is built for. Same model, same state machine, same gestures. Different output stage. The chair moved.

It is honest to say this is a first-light demonstration, not a finished product. The current build still has known issues. Right-wink commands fire too rarely and we have a fix queued. Forward commands sometimes arc slightly when the joystick module partially recovers between our fault refreshes, and a tighter refresh interval is the next change. These are refinement problems, not foundational ones.

What is foundational is now built and verified. The path from brain electrical and facial muscle activity, through a personally-trained classifier, through a confidence-thresholded state machine, through R-Net protocol injection, to motor torque at the drive wheels, is closed.