ExTraCT - Explainable trajectory corrections for language-based human-robot interaction using textual feature...

Area of Science:

• Robotics
• Human-Robot Interaction
• Artificial Intelligence

Background:

• Understanding human intent is critical for robots to align tasks with user preferences in human-robot interaction (HRI).
• Traditional methods for trajectory modification based on language corrections require extensive training and struggle with generalization across diverse scenarios.
• Existing approaches often rely on end-to-end learning, limiting adaptability and requiring large pre-trained datasets.

Purpose of the Study:

• To present ExTraCT, a modular framework for modifying robot trajectories and behavior using natural language input.
• To enable robots to adapt language corrections to new tasks, including complex motions, without additional end-to-end training.
• To offer a more explainable and versatile solution for HRI applications.

Main Methods:

• ExTraCT separates language understanding from trajectory modification, utilizing Large Language Models (LLMs) for semantic matching.
• The framework maps language corrections to predefined trajectory modification functions for robot path adjustments.
• A modular design allows adaptation to various objects, initial trajectories, and configurations.

Main Results:

• User studies in simulation and with a physical robot arm showed ExTraCT corrections were preferred in 80% of cases.
• The system demonstrated accuracy improvements over baseline methods.
• ExTraCT proved effective in complex scenarios, such as assistive feeding.

Conclusions:

• ExTraCT provides a versatile and explainable approach to interpreting language corrections in HRI.
• The modular framework overcomes limitations of traditional methods, offering adaptability across diverse applications.
• This technology facilitates robots learning human preferences and improving task alignment.