Interactively learning behavior trees from imperfect human demonstrations.

Area of Science:

• Robotics
• Artificial Intelligence
• Human-Computer Interaction

Background:

• Interactive Task Learning (ITL) enables agents to learn tasks via human instruction.
• Behavior Trees (BTs) provide a reactive and interpretable method for task encoding.
• Current ITL methods often require step-by-step input or prevent adaptation without retraining.

Purpose of the Study:

• To develop a framework for learning Behavior Trees (BTs) directly from human demonstrations.
• To enable robots to adapt learned tasks interactively without starting from scratch.
• To address limitations in current robotic ITL approaches.

Main Methods:

• Learning BTs from RGB-D video demonstrations.
• Automatic extraction of pre- and post-conditions from visual features.
• Utilizing a Backchaining approach to construct reactive BTs.
• Interactive refinement of learned BTs through a web interface to resolve failure cases.

Main Results:

• Successfully learned reactive BTs from minimal human demonstrations.
• Demonstrated the ability to automatically detect and interactively resolve failure cases at runtime.
• Validated the approach on a robotic trash disposal task with 20 participants.

Conclusions:

• The proposed framework enables efficient learning of robotic tasks from visual demonstrations.
• Interactive refinement significantly improves the robustness and adaptability of learned BTs.
• This approach advances the field of robotic ITL by allowing for dynamic task adaptation.