Leveraging Haptic Feedback to Improve Data Quality and Quantity for Deep Imitation Learning Models

Area of Science:

• Robotics
• Artificial Intelligence
• Human-Robot Interaction

Background:

• Learning from demonstration (LfD) is crucial for robot skill acquisition.
• The quality and quantity of demonstration data directly impact model performance.
• Existing teleoperation systems can be enhanced to improve data collection.

Purpose of the Study:

• To enhance a teleoperation system with real-time haptic feedback for human demonstrators.
• To evaluate the impact of haptic feedback on data throughput and quality.
• To assess the performance improvement of autonomous policies trained with haptic-enhanced data.

Main Methods:

• Implemented real-time haptic feedback in a mobile manipulator robot's teleoperation system.
• Collected door-opening demonstration data with and without haptic feedback across eight real doors.
• Trained six image-based deep imitation learning models (three with haptic data, three without).
• Evaluated autonomous door-opening policy performance.

Main Results:

• Haptic feedback improved data throughput by 6% during teleoperation.
• Autonomous policies trained with haptic-enhanced data performed 11% better on average.
• Demonstrated improved data collection efficiency and policy performance.

Conclusions:

• Real-time haptic feedback is a valuable enhancement for teleoperation data collection.
• Haptic feedback leads to higher quality data, resulting in superior autonomous robot policies.
• This method offers a practical approach to improving robot learning from demonstration.