Brain signals from fNIRS can train robots offline via RL

A team led by Julia Santaniello at Tufts University has published a paper demonstrating that brain signals captured via functional near-infrared spectroscopy (fNIRS) can effectively guide reinforcement learning (RL) for robot behavior without requiring real-time brain-computer interfaces. The study, titled "An offline approach to fNIRS-guided reinforcement learning for robot behavior," explores a human-in-the-loop RL setup where neural data is used to augment—not replace—traditional RL parameters.

The researchers tested two interaction paradigms: passive observation and active demonstration. They found that incorporating fNIRS signals into trajectory priorities and state-action Q-values improved agent learning in simulation. Importantly, the framework learns from offline data, sidestepping the need for live BCI setups. This opens the door for practical applications where real-time neural feedback is impractical or only limited datasets exist. The results are preliminary but show promise for aligning robot behavior with user preferences using only pre-recorded brain activity.