Hybrid Adaptive Brain-Computer Interface Using Artificial Intelligence for Controlling Industrial Robots and Medical Assistive Devices

Manual control is impractical in two very different contexts: for users with severe motor impairment, and in industrial settings where hands-free operation is needed. This project investigates whether a single hybrid, adaptive BCI architecture can address both — enabling hands-free, non-invasive EEG control of industrial robotic manipulators and medical assistive devices.

Can a single hybrid, adaptive AI architecture reliably decode motor-imagery EEG in real time to control both industrial robotic manipulators and medical assistive devices, while adapting to inter- and intra-session EEG non-stationarity?

The system is architected as a modular ROS 2 node graph: real-time EEG acquisition and decoding feed a hybrid adaptive AI core, which translates decoded motor-imagery intent into validated velocity and pose commands for the robot, with simulation-first verification in Gazebo before hardware deployment.

Real-time EEG acquisition uses BrainFlow for hardware-agnostic streaming, feeding an end-to-end Python pipeline built for continuous, real-time signal capture.

MNE-Python handles filtering, artifact rejection, epoching and spatial feature extraction ahead of classification.

Artifacts are removed using Independent Component Analysis (ICA) within the MNE-Python preprocessing pipeline, ahead of epoching and feature extraction.

Spatial features are extracted from preprocessed EEG epochs, feeding both the deep learning classifiers and the classical CSP-based baseline.

Motor-imagery intent is decoded using deep learning classifiers trained and evaluated in PyTorch and TensorFlow, benchmarked against classical machine learning baselines built in Scikit-learn.

CNN and LSTM architectures were designed, trained and evaluated in PyTorch/TensorFlow for motor-imagery intent decoding from EEG.

Deep learning classifiers were benchmarked against classical CSP (Common Spatial Patterns) feature extraction combined with LDA and SVM classifiers in Scikit-learn, providing a non-deep-learning baseline for comparison.

An online learning layer continuously recalibrates the decoder against inter- and intra-session EEG non-stationarity — the defining adaptive component of the system, designed to reduce per-session recalibration time without full retraining.

The robot-side control stack is a modular ROS 2 node graph that translates decoded neural intent into validated velocity and pose commands.

All robot behavior is verified in Gazebo simulation before any hardware deployment — a simulation-first approach used to de-risk real-world testing.

A shared-control arbitration layer blends decoded human intent with autonomous motion planning and collision avoidance, so low-confidence neural commands degrade safely rather than producing unintended motion.

Human-robot interaction safety is enforced through workspace limits, velocity clamping, confidence thresholding, and a watchdog-driven emergency-stop path, keeping the system fail-safe under signal dropout or decoder uncertainty.