Action Generation Adapted to Low-Level and High-Level Robot-Object Interaction States.

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Robots operate in complex, dynamic environments with changing object states.
• Effective robot task execution requires identifying actions based on object states and adapting to state changes.
• Existing methods often struggle to integrate low-level movement control with high-level decision-making for dynamic adaptation.

Purpose of the Study:

• To introduce Interaction State-based Skill Learning (IS²L), a method for robots to build skills for realistic environments.
• To enable robots to infer actions based on both low-level and high-level object states.
• To combine dynamical systems for low-level movement adaptation with affordance-based models for high-level action selection.

Main Methods:

• IS²L builds skills as Bayesian Networks that infer robot end-effector movements.
• The method transforms kinesthetic demonstrations into state representations, capturing robot-object relationships and future movements.
• Skills are learned by inferring the next movement based on current robot and object states, inspired by affordance models.

Main Results:

• The developed IS²L method successfully builds skills for tasks in realistic environments.
• The learned skills effectively utilize both high-level and low-level states for action inference and execution.
• The system demonstrated adaptability to external perturbations during task execution in experiments.

Conclusions:

• IS²L provides a unified approach to skill learning by integrating dynamical systems and affordances.
• This method enables robots to perform tasks in dynamic environments by adapting actions based on multi-level state information.
• The findings highlight the potential of IS²L for enhancing robot autonomy and adaptability in complex scenarios.