Stillwater, Oklahoma — In a major step forward for human-robot collaboration, engineers at Oklahoma State University have developed a neuroadaptive AI system that allows robots to detect human error signals directly from the brain — reacting in milliseconds to prevent disasters in high-risk environments.
The breakthrough, detailed in research from OSU’s iHuman Lab and published earlier this month, uses non-invasive electroencephalogram (EEG) caps to capture Error-Related Potentials (ErrPs). These are specific electrical patterns generated in the brain’s anterior cingulate cortex the instant a person recognizes a mistake — faster than any physical reaction can occur.
“ErrPs are specific electrical patterns generated by your brain, specifically the anterior cingulate cortex, the moment you recognize a mistake,” explained lead researcher Dr. Hemanth Manjunatha, assistant professor in the School of Mechanical and Aerospace Engineering. “The fascinating part is that your brain reacts to an error faster than you can physically move your hand to fix it.”
In teleoperation scenarios — where humans remotely control robots in dangerous settings such as nuclear decommissioning, deep-sea inspections, or handling space debris — operators often face mental fatigue and delayed responses. The new system gives robots an “early warning” by interpreting these instinctive “Oh no!” brain signals and triggering immediate safety actions.
Using a combination of self-supervised learning and Signal Temporal Logic (STL), the AI first builds a general model of brain patterns and then fine-tunes it for each individual user in seconds — similar to how a smartphone learns to recognize a new face. Once an ErrP is detected, STL mathematically enforces safe robot behaviors: controlled braking, full stop, or handing control back to the human.
All processing runs in real time on NVIDIA Isaac Lab and Isaac ROS platforms powered by NVIDIA RTX PRO 6000 GPUs, ensuring decisions happen faster than an emergency button could be pressed.
Dr. Manjunatha emphasized that full robot autonomy remains too risky in unpredictable real-world conditions. “In high-stakes environments, like decommissioning a nuclear site or performing deep-sea inspections, we can’t yet turn the keys over entirely to a robot. The world is too unpredictable,” he said. “By detecting these ErrPs, we aren’t just reading brain activity; we are capturing the human’s instinctive moment.”
The technology also holds promise for healthcare. Future applications could include smart prosthetics or exoskeletons that sense when a user feels a movement is wrong and self-correct automatically, making the device feel like a natural extension of the body.
The entire system — including code, models, and datasets — will be released as open source to accelerate adoption in rehabilitation and mobility assistance for people with disabilities. Graduate and undergraduate students in the iHuman Lab, including Ph.D. student Elahe Oveisi and master’s student Bennett Dogbey, played key roles in development and testing.
