Artificial Intelligence-Driven Brain-Computer Interfaces based on Electroencephalography towards Robotic Arm Control

: (Alternative Format Thesis)

Abstract

Electroencephalography (EEG)-based brain–computer interfaces (BCIs) have emerged as a promising non-invasive pathway for robotic arm control. By decoding neural activity directly from scalp recordings, EEG BCIs circumvent peripheral neuromuscular pathways and offer a comparatively low-cost, portable, and safe alternative to invasive approaches. Crucially, recent progress in artificial intelligence (AI) has begun to transform EEG BCIs from rigid command interfaces into machine-intelligent control systems capable of robust intention inference and context-aware assistance.

This PhD project systematically investigates key challenges in AI-enabled EEG-driven robotic arm control, spanning preprocessing, neural decoding, user adaptation, and human–robot collaboration. The first strand of work examines how preprocessing choices shape the input space presented to downstream machine learning decoders in motor imagery (MI)-based BCIs. A range of methods—including independent component analysis, surface Laplacian, bandpass filtering, and baseline correction—are rigorously compared across multiple datasets and classifiers. The results show that comparatively simple strategies such as baseline correction and bandpass filtering consistently enhance classification performance, while combining surface Laplacian with spatially sensitive learning algorithms yields additional gains. These findings provide concrete guidance for building efficient, deployment-oriented AI pipelines for online robotic arm control, where computational budget, latency, and robustness are essential.

Building on this foundation, the second strand develops \textbf{MCSRTNet}, a multiscale clustered spatial residual network tailored to EEG-based MI decoding. By integrating phase-locking-value (PLV)-guided frequency-specific spatial grouping, spatio--spectral channel augmentation with residual learning, and lightweight log-variance-based temporal modeling, MCSRTNet enables structured feature learning that better captures discriminative patterns in noisy, low-SNR EEG signals. Comparative experiments on benchmark datasets demonstrate that MCSRTNet outperforms representative state-of-the-art architectures and provides a more robust and interpretable framework for MI decoding, thereby offering stronger support for stable downstream decision-making and robotic control.

The third strand addresses a major limitation of current BCIs: their vulnerability to fluctuations in users’ mental states. A dual-paradigm experimental protocol is introduced that combines active BCI (aBCI) for MI decoding with passive BCI (pBCI) for frustration recognition. Real-time affective information is incorporated into AI-based adaptive decoding pipelines that recalibrate to the user’s state, effectively treating cognitive–emotional context as an additional signal for machine intelligence. Experimental results show that this combined aBCI–pBCI approach improves decoding accuracy and reduces variability relative to conventional static models, thereby advancing toward user-adaptive intelligent BCIs that learn under non-stationarity.

The final strand proposes a mental state-regulated shared autonomy framework using brain-computer interface and machine intelligence for robotic arm control. This framework integrates MI-based intention decoding (aBCI) and real-time mental state recognition (pBCI) into an AI-driven inference model that feeds a dynamic arbitration policy between human and robot. Control authority is continuously adjusted according to both inferred motor intent and affective state, enabling an intelligent autonomy layer that supports the user while preserving agency. In real-world robotic arm experiments, this emotionally informed shared-autonomy system reduces subjective workload, improves perceived agency and sense of control, and increases task success rates. These findings demonstrate the feasibility and utility of incorporating affect-aware AI decision-making into assistive robotic control.

In summary, this PhD project advances EEG-based robotic arm control along four dimensions: (i) systematic evaluation of preprocessing pipelines, (ii) development of a dedicated deep learning decoder for MI, (iii) incorporation of affective-state modeling for adaptive calibration, and (iv) realization of a collaborative control framework that unites intention and emotion in shared autonomy. Together, these contributions show that EEG signals, when combined with advanced machine learning and human-state modeling, can support robotic arm systems that are more reliable, adaptive, and user-centred, and that have the potential to enhance autonomy and quality of life for individuals with severe motor disabilities.

Date of Award	24 Jun 2026
Original language	English
Awarding Institution	University of Bath
Supervisor	Benjamin Metcalfe (Supervisor) & Dingguo Zhang (Supervisor)