Dynamic Input Deep Learning Control of Artificial Avatars in a Multi-Agent Joint Motor Task

Area of Science:

• Robotics
• Human-Robot Interaction
• Artificial Intelligence
• Motor Control

Background:

• Real-world scenarios increasingly require seamless human-robot coordination for shared objectives.
• Existing research predominantly focuses on dyadic (one-on-one) human-robot interactions, leaving multi-agent scenarios with numerous humans under-explored.
• Synthesizing autonomous agents capable of fluidly interacting within human ensembles is a significant challenge.

Purpose of the Study:

• To develop an autonomous artificial agent capable of performing oscillatory joint tasks with human groups.
• To ensure the artificial agent exhibits human-like kinematic features during interaction.
• To create a flexible architecture adaptable to human groups of varying sizes.

Main Methods:

• A deep reinforcement learning (DRL) architecture was designed for agent control.
• The paradigmatic multi-agent mirror game, an oscillatory motor task, was employed to test coordination.
• The DRL approach was optimized for adaptable interaction with diverse human group sizes.

Main Results:

• The proposed DRL architecture successfully enabled an artificial agent to coordinate movements in a multi-agent mirror game.
• The agent demonstrated the ability to interact with human ensembles of different sizes.
• The system showed potential for exhibiting desired human kinematic features during joint tasks.

Conclusions:

• Deep reinforcement learning offers a viable solution for synthesizing autonomous agents in complex human-robot group tasks.
• The developed architecture provides a flexible framework for human-robot coordination in multi-agent settings.
• This research advances the field of human-robot interaction by addressing multi-agent coordination challenges.