Your Brain Knows When the Robot Messes Up: EEG Error Signals Could Transform Human-Robot Teamwork

Imagine working alongside a robot that can sense, from your brainwaves alone, the exact moment it makes a mistake. Long before you raise your hand, shout a correction, or reach for the emergency stop, your nervous system has already registered the error, producing a characteristic burst of electrical activity on your scalp. Researchers have now systematically reviewed how this neural signal, known as an error-related potential or ErrP, can be harnessed to make human-robot collaboration safer, smarter, and more intuitive. The findings suggest that the human brain may become one of the most valuable sensors in the factory, the hospital, and the home.

The review, published in Frontiers in Neuroergonomics by David Achanccaray of the Fédération ENAC ISAE-SUPAERO ONERA at the University of Toulouse, Aurélie Clodic of LAAS-CNRS, and Raphaëlle N. Roy of the University of Toulouse, examined the small but growing body of research at the intersection of neuroscience and robotics. The team applied the PRISMA protocol, a standardized framework for systematic reviews, searching four major databases including Web of Science, IEEE Xplore, Scopus, and Google Scholar in October 2025. From an initial pool of roughly forty publications, thirteen studies survived the screening and eligibility stages. All of them had been published within the last nine years, and notably, every single one was conducted in a laboratory environment rather than in the field.

ErrPs are not new to neuroscience. They were first characterized in the early 1990s, when researchers identified a negative deflection in electrical brain activity, called error-related negativity, appearing over the fronto-central cortex roughly 80 milliseconds after a person makes an error in a choice task. This is followed by a larger positive component over the centro-parietal cortex between 200 and 500 milliseconds. Error events also modulate brain rhythms: theta-band activity between 4 and 8 hertz increases, while beta-band activity between 12 and 30 hertz decreases. Connectivity analyses have pointed to the anterior cingulate cortex influencing the prefrontal cortex during error processing. A related signal, the feedback-related negativity, appears 200 to 300 milliseconds after a person receives feedback indicating wrong performance. Crucially, these potentials arise unwittingly, meaning the brain flags errors automatically, even when the mistake is made by a robot, another person, or an artificial intelligence system rather than by the observer.

To capture these signals, researchers use a passive brain-computer interface, or BCI. Unlike active BCIs, where users deliberately modulate their brain activity to control a device, passive BCIs monitor brain activity in the background while the user engages normally with the task. In the context of human-robot collaboration, where humans and robots share goals and workspaces, coordinating through physical forces or through communication channels such as gestures, voice, gaze, and facial expressions, a passive BCI can detect the user’s implicit evaluation of the robot’s behavior. When the system decodes an ErrP, it can trigger feedback, initiate a corrective action, or feed the signal directly into the robot’s control algorithm, allowing the machine to adapt to the human in real time.

The thirteen studies fell into three broad categories: adaptation, control, and neuroergonomics. In the adaptation studies, the most striking results came from robot learning experiments. In one line of work, humans taught a robot to associate gestures with actions, and when the robot executed the wrong action, the resulting ErrP served as a reward signal in a reinforcement learning algorithm, updating the robot’s behavior policy. Detection accuracies reached 87.2 percent from the human perspective and 86.9 percent from the robot perspective, and the researchers found a strong correlation, with correlation coefficients between 0.869 and 0.965, between how well the ErrP was detected and how well the robot learned. In a gaming scenario, where a human guessed a robot’s gaze pattern among three objects, ErrP-triggered updates to both human and robot behavior policies improved the percentage of correct guesses from 33 percent to between 70 and 90 percent within just 10 to 40 trials.

Supervision and safety applications showed particular promise for industrial settings. In one study, an autonomous robot selected targets for a drilling operation, and when it chose incorrectly, the detected ErrP triggered a correction sequence in which the human supervisor’s hand gestures, classified from electromyographic signals, told the robot what to do instead. This hybrid EEG-EMG framework achieved 97.3 percent correct targets, even though the ErrP detection itself averaged only around 54 percent with a small participant pool. Another study simulated emergencies in an assembly task, including crush, drop, and unexpected movement events, and found that an optimal time window of 250 milliseconds could be used to process EEG signals for a robot stop interface. The authors noted, however, that their participants lacked real human-robot interaction experience, which likely shaped the neural responses compared with professional robot operators.

Control applications pushed ErrPs deeper into the robot’s decision loop. In a virtual smart wheelchair scenario, the control algorithm recalculated its path planning whenever an ErrP signaled that the user perceived the wheelchair approaching an obstacle, with detection accuracy reaching 78.6 percent. A follow-up comparison against steady-state visually evoked potentials, a different type of brain signal requiring deliberate attention to flickering stimuli, found that the ErrP approach achieved higher task achievement but lower user engagement. In another experiment, a robotic arm moving along a one-dimensional grid received ErrP-derived feedback during reinforcement learning training, which reduced the variability of the learning process and led to convergence after roughly nine iterations. Even in complex control scenarios involving singularity avoidance, where robot dynamics become unpredictable, detectable ErrPs emerged, and participants preferred the strategy that produced greater robot predictability, reflected in lower power spectral density of the error signals.

The neuroergonomics studies revealed a critical caveat: mental state matters. In one experiment, participants controlled a robot arm end-effector while simultaneously performing mental calculations of varying difficulty. When the robot stopped unexpectedly in 33 percent of trials, the researchers found that high mental workload dampened the ErrP amplitude, indicating reduced error awareness. Theta-band power increased in the frontal cortex, with trends toward increased alpha and beta power in the central cortex after unexpected stops. A second study explored boredom during a collaborative maze exploration task, where the robot consulted the participant about direction at intersections and deliberately changed course in 30 percent of trials. As boredom set in during the second half of the task, ErrP amplitude and theta-band power declined, suggesting that disengagement blunts the very neural signal these systems depend on.

Across all thirteen studies, detection accuracies ranged from 54 to 87.2 percent, and in most experiments the robot’s mistakes were programmed to occur randomly in 20 to 35 percent of trials. The authors argue that while provoking errors is experimentally convenient, naturally occurring errors during genuine robot learning or environmental interaction would better reflect real-world conditions. They envision contactless interaction in exploration, search, or assembly tasks as the most promising scenarios for deploying ErrP-based passive BCIs, particularly within human-centered AI frameworks where the system continuously updates robot behavior policies based on decoded brain signals. Safety, they emphasize, will require wireless EEG systems that remain securely fixed to prevent interference with the shared workspace.

The review acknowledges that the field remains young, with all published work confined to laboratories and small samples. Yet the consistency of the findings, from gesture learning to wheelchair navigation to assembly supervision, suggests that ErrPs offer a genuine channel through which robots can perceive human judgment without a single word or gesture. As detection algorithms improve and wearable EEG technology matures, the line between human oversight and machine autonomy may blur in ways that make collaboration not just more efficient, but fundamentally safer. The brain, it turns out, has been flagging robot errors all along. Now the robots are finally learning to listen.

Subject of Research: Detection of error-related potentials from EEG for enhancing human-robot collaboration

Article Title: Error-related potentials detection to enhance human-robot collaboration: a mini review

Article References: Achanccaray, D., Clodic, A., & Roy, R. N. (2026). Error-related potentials detection to enhance human-robot collaboration: a mini review. Frontiers in Neuroergonomics, 7, Article 1769098. https://doi.org/10.3389/fnrgo.2026.1769098

Image Credits: AI Generated

DOI: 10.3389/fnrgo.2026.1769098

Keywords: error-related potentials, human-robot collaboration, brain-computer interface, EEG, passive BCI, robot learning, reinforcement learning, neuroergonomics, mental workload, adaptive control, human-robot interaction, Frontiers in Neuroergonomics