Intrinsic interactive reinforcement learning - Using error-related potentials for real world human-robot interaction.

Area of Science:

• Robotics
• Neuroscience
• Machine Learning

Background:

• Reinforcement learning (RL) in robotics typically relies on explicit reward functions, which can be burdensome for human operators.
• Continuous explicit feedback is demanding and tiresome, highlighting the need for implicit feedback mechanisms.

Purpose of the Study:

• To develop and validate an implicit human feedback system for robot reinforcement learning using electroencephalogram (EEG) signals.
• To enable robots to learn optimal behavioral strategies in dynamic environments by leveraging intrinsically generated human feedback.

Main Methods:

• Utilized error-related potentials (ErrPs), an event-related activity in human EEG, as an implicit reward signal for RL.
• Developed an online single-trial ErrP detection system with high balanced accuracy (bACC).
• Validated the approach in both simulated and real-world robot learning scenarios involving gesture recognition and action mapping.

Main Results:

• Achieved high online ErrP detection accuracy (91% bACC in simulation, 90% bACC in real-world).
• Demonstrated successful parallel learning of gesture recognition and human-robot action mapping.
• Confirmed the robot's ability to correctly learn gesture-action mappings based on implicit EEG feedback.

Conclusions:

• Intrinsically generated EEG-based human feedback can effectively improve gesture-based robot control in human-robot interaction.
• The developed approach, termed intrinsic interactive RL, offers a less demanding alternative to explicit feedback for robot learning.
• This work demonstrates the potential of leveraging neurophysiological signals for more intuitive and efficient human-robot collaboration.