Shared autonomy between human electroencephalography and TD3 deep reinforcement learning: A multi-agent copilot...

Area of Science:

• Neuroscience
• Artificial Intelligence
• Robotics

Background:

• Deep reinforcement learning (RL) enables autonomous agents.
• Brain-computer interfaces (BCI) decode human brain signals.
• Integrating RL and BCI can improve autonomous system performance and human intervention.

Purpose of the Study:

• To propose a novel integration technique between deep RL and BCI.
• To enhance human interventions in autonomous systems.
• To improve BCI performance by considering environmental factors.

Main Methods:

• Developed a copilot control scheme (Co-FB) with shared autonomy between human (EEG) and RL (TD3) agents.
• Utilized electroencephalography (EEG) for human action command decoding.
• Employed twin delayed deep deterministic policy gradient (TD3) for RL agent actions.
• Introduced a disparity index to evaluate conflicting agent decisions.

Main Results:

• The Co-FB model significantly outperformed individual EEG (EEG-NB) and TD3 control.
• Co-FB achieved higher target-approaching scores, lower failure rates, and reduced human workload.
• Shifting control authority to the TD3 agent improved performance during suboptimal BCI decoding.

Conclusions:

• The copilot system effectively manages complex environments.
• Integrating environmental factors enhances BCI performance.
• This approach offers improved human-AI collaboration in autonomous systems.