AutoNeuro

AutoNeuro — Imagined Speech-to-Intent BCI Inspiration Currently, humans interface with AI via external modalities such as typing and text-to-speech. A more seamless method of human-AI connection would be to communicate directly from the brain to the AI. However, this hinges on accurate internal speech decoding, which has been a long standing, unsolved problem to do so accurately in the field of Neurotech. Noninvasive approaches, especially, suffer from low SNR. Is there a way to implement new frontier tech, such as Zuno & AutoResearch, to create a self-optimizing model with competent accuracy compared to conventional methods?

What it does Taking inspiration from Google Robotics pioneering the VLA model architecture, we attempted to create something similar in that we wanted a model that could output action tokens that represent semantic meaning given EEG recordings of internal speech.

How we built it Signal ingestion: EEG (32-channel, 250Hz) preprocessed via ZUNA

Encoder: EEG encoder producing 512 Dimensional latent vectors

Projector: 3-layer MLP mapping 1024D brain vectors Classifier backend: MLP command classifier on WAY-EEG-GAL grasp/lift EEG epochs for discrete action decoding Robot arm: FastAPI + WebSocket simulator receives action tokens and animates arm movements UI: Gradio dashboard + agentic optimization loop running iterative self-improvement passes Datasets: WAY-EEG-GAL (grasp/lift EEG), ZuCo 2.0, OpenNeuro ds004514

Challenges we ran into

Signal quality gating — noisy or low-confidence epochs needed abstention logic rather than bad predictions

The biggest problem we ran into was having to switch tracks on the second day. ]On the technical side, we struggled to get computers to execute our ideas and test MVPs because all our team members are on consumer grade hardware that could not handle loading the larger models, much fine tuning or training a frontier model. Furthermore, incorporating the fNIR data with the EEG data was challenging because of the delay between the near instant EEG voltage changes and the slower metabolic changes. Also a lesson we learned was getting too excited and making architecture decisions before understanding the current state of the field. Initially we wanted to use NueGPT but after debating on how we can use it, we discovered that the frontier had actually moved on to larger foundational models such as Brain-OF that natively handles multimodal data from nearly every brain signal imaginable.

Accomplishments that we're proud of

We were successfully able to implement a proof of concept wherein we created a self optimizing model that can decode EEG signals into action tokens that move a simulated robotic arm. While this used motor intent, the concept is highly transferable as semantic and motor intent use similar neural pathways when it comes to imagining action related words like the ones we used. Hence, we can utilize the same pipeline for semantic intent by just retraining with semantic data.

Although we were not able to accomplish all that we wanted to, we were happy with frontier tech. Developing the models via AutoResearch, agentic research and an end to end model training pipeline.

What we learned

Neuroscience and ML have a steep integration curve — domain-specific preprocessing (MNE, artifact rejection, bandpass filtering) is as important as the model architecture Multimodal biological signals require careful synchronization design, not just concatenation LLMs can be grounded with non-text modalities if the projection layer is trained carefully Agentic loops are powerful for rapid iteration but need robust failure handling

What's next for AutoNeuro

Train the projector on larger paired EEG-text datasets for higher decoding accuracy Expand action token vocabulary beyond grasp/lift to support richer motor intent Explore real hardware integration beyond the simulated arm Optimize latency for sub-second closed-loop feedback [PLACEHOLDER: Any plans to pursue this post-hackathon? Commercialization, research paper, open-source release?]

Built With

• autoresearch
• zuno