NEURO-LOOP (Coming Soon)
Brain-computer interfaces for human-in-the-loop RL
NEURO-LOOP is a reinforcement learning framework that integrates implicit neural feedback into robotic and artificial agent policy learning. In this iteration, we explore using a hybrid fNIRS-EEG system to train the agent in a real-time loop. Further, we classify multiple brain states and signals to further improve the agent's performance.
Overview
Previous work used passive neural activity as a supervisory signal to enable agents to adapt its behavior to a user’s expectations as they naturally observe the agent’s behavior. Current work explores using a hybrid fNIRS-EEG system classify performance, preference, error and attentional states to further improve a robotic agent's performance.
We (will) release a dataset of fNIRS and EEG data collected from the human-robot interaction, consisting of the learning statistics of the robot, and the brain data of the human. Finally, we apply this framework in real-time to assess performance.
Focus
This work focuses on how to reliably use various biological signals to update the policy of a robotic agent by using real-time brain-computer interface (BCI) data.
(Currently under development)