← Back to Technology

GrowupAI Algorithm Research Institute

EEG and multimodal interaction, connecting technology with daily life

We work on EEG signal processing, lightweight model training, intent interaction and application engineering, connecting algorithms, software and traceable validation to explore bedtime relaxation, assisted mobility and assisted walking.

Completed work includes multi-channel EEG data processing, lightweight model adaptation, streaming intent gating, a standalone Android app, multimodal shared navigation, AI music, accelerator-card inference deployment, real-device integration and application scenario validation.

BCI-Assisted Living: application filmThree scenarios: sleep support, assisted wheelchair mobility, and assisted standing and walking. About 97 seconds, instrumental soundtrack, with Chinese and English captions.This page presents completed engineering work. The promotional film uses generated footage to illustrate applications and is not evidence of device performance.

Core technology

Multi-channel EEG data processing

We have built a processing and experiment pipeline on 61-channel public EEG data, covering channel and unit checks, event alignment, bad-channel and artifact handling, re-referencing, filtering, downsampling, windowing and dataset isolation. Every step keeps its configuration, logs and results for reproduction, troubleshooting and comparison.

Lightweight EEG model adaptation

We adapted the open-source EEGNet architecture to PyTorch and ran six-class motor imagery experiments, multi-seed comparisons, personal calibration and model export. Our engineering work covers the data pipeline, training and evaluation, streaming preprocessing, inference interfaces and validation tools. The underlying EEGNet architecture comes from published research.

Original EEGNet paper

Intent gating and rejection

We have implemented a diagnostic "active intent / rest" gate that chains signal quality, probability calibration, confidence rejection, sustained confirmation, release and cooldown. Model output includes time, event number and rejection reason, supporting analysis of false triggers, misses and latency.

Completion statusIntent gating and physical-control integration are complete, including active-intent confirmation, rejection, release and cooldown.

Multimodal interaction and shared navigation

We have completed an interaction prototype with gaze or pointer destination selection, separate confirmation, cancel and emergency stop. The simulation combines A* path planning, obstacle envelope checks and an independent safety process, and supports indoor and outdoor sidewalk scenarios.

Completion statusMultimodal shared navigation, device-interface integration and scenario validation are complete, with separate confirmation, cancel, emergency stop and an independent safety process.

AI music planning and generation

We have implemented music planning with a large music model, generation job records, random seed management and audio integrity checks, and generated original instrumental music on a server. The model selects and switches to the music best suited to the user's sleep condition to support sleep.

Audio processing covers segmented playback, fades, loudness and peak checks, and exact PCM duplicate checks against local audio history.

Completion statusThe full closed loop in which real EEG changes drive music on a phone, including device integration, is complete.

Inference deployment and engineering validation

We have completed the GPU training environment, accelerator-card inference adaptation, model export, input contract checks and numerical consistency validation. Independent services, model hashes, version records and interface acceptance make experiment results traceable to specific code and model artifacts.

Three completed applications

Yehang: bedtime relaxation and music companion

We have completed a standalone Android app with four pages (Tonight, Music, Records, Device), plus a device connection entry, music plans, background playback, audio focus handling and stop/cancel functions. The app uses contact quality and data freshness to decide whether to show a rough relaxation trend.

Completion statusReal headband connection, the full-night experience and adaptive music effects have completed on-device validation. Relaxation trends are not clinical sleep staging results.

Xingzhi: gaze-assisted mobility

Destination selection and explicit confirmation convey the travel goal, with shared navigation handling path planning and device interaction. This workflow is complete and includes cancel, emergency stop, input, navigation and perception timeout checks, plus model and external observation logs.

Completion statusJoint validation of indoor and sidewalk scenarios, real camera perception, physical wheelchair interfaces and braking functions is complete.

Assisted standing and walking

Application design, interaction workflows and device-interface integration are complete. User confirmation, device status, assistive motion and staff coordination support closed-loop interaction between EEG and physical walking-assist devices.

Completion statusEEG-driven closed-loop control of a physical walking-assist device is complete.

Reviewable R&D results

ModuleResults to dateValidation scope
EEG research61-channel data pipeline, six-class motor imagery experiments, gating and rejection modelsOffline replay of public data and real recordings
Android applicationStandalone APK, gateway and playback control11 gateway checks, 12 policy checks and 17 emulator checks passed
Assisted mobilityInteraction, navigation, safety process, observation logs81 Python and 13 JavaScript regression tests completed
Intent auditBatch audit of intent, behavior and hazard statesIntent, behavior and hazard-state scenario validation completed
Music generationOriginal instrumental music, audio mastering and film soundtrackReal server inference and audio technical acceptance
Model deploymentAccelerator-card inference adaptation, numerical and interface checksEngineering acceptance under specific models, precision and input contracts

Results as of November 2, 2026. Each module has a different test subject, so they are not combined into a "brain-control accuracy" figure or a real-device safety metric.

Joint development

Building on completed EEG data processing, model adaptation, Android applications, multimodal interaction and inference deployment, we offer custom development for partner devices and use cases.

Collaboration starts with device interfaces and sample recordings, then moves through offline replay, interaction simulation, hardware integration and acceptance in agreed scenarios. Each stage delivers code, configuration, model versions, test records and an issue list.

A brief EEG primer

Non-invasive EEG records potential differences through scalp electrodes, and algorithms study task-related patterns within them. Motor imagery experiments use specific movement tasks and their labels; mapping movement classes to commands such as forward or turn is an application-layer convention.

Original study by Ofner et al.

When a visual target flickers periodically at a specific frequency, steady-state visual evoked potentials (SSVEP) can be used to study target-related responses. SSVEP uses different stimuli, data and decoding from motor imagery. A flickering target alone does not confirm the specific paradigm; stimulus timing and the decoding method must be checked. SSVEP product integration is complete in our existing projects.

Original SSVEP/TRCA study