Error-related potential-based shared autonomy via deep recurrent reinforcement learning.

Area of Science:

• Robotics
• Neuroscience
• Human-Computer Interaction

Background:

• Error-related potentials (ErrPs) are brain signals evoked by observing unexpected events, offering a non-explicit feedback channel for human-robot interaction (HRI).
• Traditional Brain-Computer Interfaces (BCIs) often require continuous operator commands, limiting their application in dynamic HRI scenarios.
• ErrP-based BCIs present a promising alternative by leveraging implicit human feedback to improve robot performance.

Purpose of the Study:

• To propose and evaluate a novel shared autonomy model for human-robot interaction utilizing ErrP-based BCIs.
• To integrate ErrP signals as observations within a shared autonomy framework to enhance robot decision-making.
• To address the inherent uncertainty in ErrP signals using advanced machine learning techniques.

Main Methods:

• Formulated the shared autonomy problem as a partially observable Markov decision process (POMDP).
• Employed a recurrent neural network-based actor-critic model to process and interpret ErrP signals.
• Evaluated the framework in simulated and real-world human-in-the-loop robot navigation tasks.

Main Results:

• The ErrP-based shared autonomy model significantly improved navigation task efficiency.
• In simulations with 70% ErrP accuracy, task completion time was reduced by 14.1% compared to no ErrP feedback.
• With real users, the autonomous robot completed navigation tasks 14.9% faster.

Conclusions:

• Shared autonomy, powered by deep recurrent reinforcement learning and ErrP-based BCIs, is an effective strategy for managing uncertain human feedback in complex HRI.
• The proposed model demonstrates the potential of leveraging implicit human neural signals to enhance autonomous system performance.
• This research opens avenues for more intuitive and efficient human-robot collaboration in dynamic environments.