University of Twente Student Theses
Few-Shot Imitation Learning of Motion- and Contact-Based Tasks using Dynamic Movement Primitives
Bolding, F.K.H. (2024) Few-Shot Imitation Learning of Motion- and Contact-Based Tasks using Dynamic Movement Primitives.
|
PDF
7MB |
| Abstract: | Traditionally, robots have primarily been deployed in industrial settings, where they carry out repetitive, pre-programmed tasks with high precision. Robots are increasingly diversifying into sectors beyond manufacturing, such as healthcare and agriculture. However, operating in dynamic, human-populated environments requires adaptability and compliance. Imitation learning simplifies the reprogramming of robots by enabling operators to demonstrate desired behavior. By encoding the demonstrations in dynamic movement primitives (DMPs), task reproductions can be generalized to adapt to changes in the environment. This paper proposes a geometry-aware LfD framework based on DMPs that allows for generalization, variable compliance, and the reproduction of motion-based and contact-based tasks. The variable compliance is driven by the variability between the demonstrations, pursuing an automatic balance between desired compliance and tracking performance. Demonstrations, which include motion and force data, are synchronized and processed using a novel weighted rolling window approach that finds their mean and standard deviations. The mean demonstration is learned with the DMP. The hybrid force-impedance controller enables simultaneous reproduction of motion and force perpendicular directions. The framework was validated on real-world tasks, demonstrating its practicality. In addition to the proposed framework, contributions include recommended corrections when encoding quaternion trajectories using DMPs to prevent singularities and artefacts. |
| Item Type: | Essay (Master) |
| Faculty: | ET: Engineering Technology |
| Programme: | Biomedical Engineering MSc (66226) |
| Link to this item: | https://purl.utwente.nl/essays/103983 |
| Export this item as: | BibTeX EndNote HTML Citation Reference Manager |
Show download statistics for this publication
Show download statistics for this publicationRepository Staff Only: item control page