Colloquia

Samenvatting

As robotic systems become more integrated into dynamic environments, enabling non-expert users to intuitively program robots remains a challenge. Learning from Demonstration (LfD) offers a promising solution by allowing robots to acquire skills through human demonstrations instead of explicit programming. Effectively storing and exploiting past user demonstrations is a fundamental challenge. A solution to this challenge is the organization of demonstrations in a skill library. However, this requires robust segmentation, classification, and novelty detection techniques to ensure effective learning and reusability.

This thesis presents a modular framework for skill recognition, segmentation, and novelty detection. The framework enables robots to identify and classify demonstrated motions by matching them to skills in a predefined skill library, ensuring that non-expert users can intuitively program robotic tasks. The system incorporates time series classification using Detach-ROCKET, change point detection for motion segmentation, and novelty detection to identify previously unseen skills. User confirmation is integrated into the process to improve robustness and ensure reliability despite imperfections in human demonstrations.

The framework is evaluated using position data of motion demonstrations recorded with the Franka Research 3 robotic arm via kinesthetic teaching. Results indicate that Detach-ROCKET with MINIROCKET kernels provides high classification accuracy while maintaining computational efficiency. Segmentation methods, including Pruned Exact Linear Time (PELT) segmentation and binary segmentation, exhibit robust segmentation. Additionally, novelty detection using Local Outlier Factor (LOF) effectively identifies new skills, enabling the expansion of the skill library.

By improving skill recognition and segmentation, this framework enhances the usability of LfD, making robot programming more accessible to non-experts. The relevance of this thesis extends beyond technical challenges. It also plays a role in making robotics more accessible in industries facing labor shortages. Future work could explore integrating additional contextual information and alternative teaching methods to further refine the learning process.