Robot Skill Learning Using Haptic Data and Conditional Variational Autoencoder

The patent uses haptic devices to capture human operator actions and creates digital copies of these actions as training data for the robot. This copying approach allows the robot to learn from realistic human demonstrations without requiring physical presence, bridging the simulation-to-reality gap by using actual human interaction data rather than simulated data.

The haptic device serves as an intermediary between the human operator and the robot learning system. It captures physical interaction forces and movements, translating them into actionable training data. This intermediary enables the collection of realistic interaction data while maintaining a controlled data collection process, resolving the contradiction between simulation control and real-world applicability.

The system uses unsupervised learning algorithms that automatically process and extract skills from haptic data without requiring manual labeling. The robot skill learning model self-organizes the captured action data into meaningful skills and policies, eliminating the time-consuming manual annotation process while maintaining high learning accuracy.

The haptic device captures comprehensive action data during normal operation before any learning process begins. This preliminary data collection in the form of unlabeled haptic traces provides rich training material that can be processed later through automated algorithms, avoiding the need for time-consuming labeled data collection during the learning phase.

The patent extracts essential skill patterns from complex haptic data using unsupervised learning algorithms. Instead of processing all raw data details, the system identifies and extracts meaningful skill representations and action patterns, reducing computational complexity while preserving the essential precision needed for accurate robot manipulation.

The complex manipulative actions are segmented into discrete skills and sub-skills through the learning process. The haptic data is divided into meaningful action units that can be independently learned and combined, reducing overall processing complexity while maintaining precision for each individual skill component.