In their article “Design of a brain-machine interface for reducing false activations of a lower-limb exoskeleton based on error related potential” Soriano-Segura and colleagues address one of the classic bottlenecks of motor imagery-based brain-machine interfaces for controlling exoskeletons: false activations. Instead of relying solely on the IM classifier, the team proposes integrating error-related potentials (ErrPs) as an additional layer of supervision for exoskeleton control. To do this, they design a specific protocol to evoke ErrPs associated with incorrect activations of the lower-limb exoskeleton and develop an iterative parameter selection method that optimizes the detection of these potentials in EEG. The result is a BMI capable of self-correcting erroneous commands in near real time, with a direct impact on the safety and usability of the system during assisted walking with Exo-H3.
The most recent work, “Characterization of error-related potentials during the command of a lower-limb exoskeleton based on deep learning” builds on this foundation and takes a step further in two key directions: detailed characterization of ErrP in exoskeleton walking tasks and improvement of its detection through deep learning. In this case, the group uses the EEG-Inception network as a convolutional model to identify ErrP during the initiation and termination of assisted walking, first in static conditions and then in motion. In addition, they introduce data augmentation techniques to address the inherent imbalance of ErrP datasets. The results show an improvement in accuracy and robustness compared to traditional classifiers, confirming that the combination of deep models and specific evocation protocols allows for better exploitation of the brain signal associated with error during exoskeleton use.
Taken together, these two studies advance toward an error-related potentials exoskeleton control architecture in which the user not only generates walking commands through BMI, but also implicitly “evaluates” and corrects the robot’s behavior through their own neural responses. The integration of ErrP, first as a mechanism for reducing false activations and then as a signal exploited with deep learning models, illustrates a clear path to increasing the reliability of brain-machine interfaces in rehabilitation. Beyond the quantitative improvement in classification metrics, the main contribution is conceptual: placing continuous monitoring of human-robot interaction at the center of control design, an approach that is likely to extend to future generations of neurorehabilitation systems and robotic exoskeletons