EUVIMEDEUVIMED entdecken
Crossref

Learning a Unified Dynamic System Model Across Diverse Robotic Demonstration Tasks

Zhehao Jin, Weiyong Si, Xu Ran, Chenguang Yang, Chen Lv

IEEE Transactions on Automation Science and Engineering · 2026 · 23 · 14290-14305

DOI: 10.1109/tase.2026.3718280 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Transactions on Automation Science and Engineering, 2026. Verfasst von Zhehao Jin und 4 weiteren Autor:innen. Die Arbeit nennt 63 Literaturangaben. Gefördert von: Nanyang Technological University, Singapore, Ministry of Education, Singapore.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Learning dynamic systems (DSs) to encode trajectories has been shown as an effective way for robotic learning from demonstration. However, current DS learning-based approaches cannot generalize across different demonstration tasks due to their simple model formulations. The main reason that limits the flexibility of DS models is the stability guarantee, which is a necessary property to ensure the safety. In this paper, we solve this problem by designing a task-parameterized energy function (TP-EF), which can adapt across different demonstration tasks and serve as the stability certificate. Specifically, except for the system state, the TP-EF additionally takes a task parameter as input, which allows it to generalize across different tasks. We show that the designed TP-EF has a unique minimum for any fixed task parameter, which makes it a proper stability certificate. A TP-DS is then formulated and learned based on the TP-EF, allowing it to be a unified model across different demonstrations. Both simulation and experiments are conducted to validate the strong generalization ability of the proposed TP-DS. Note to Practitioners—In current industrial scenarios, robots always complete tasks by reproducing human demonstrations, which is known as programming by demonstration. Comparatively, learning from demonstration (LfD) further provides robots with generalization ability, such as allowing generalization across different start and goal postures. However, the generalization ability of current LfD approaches is quite weak. Existing LfD approaches cannot principally generalize across different tasks due to their simple model structures. For example, one LfD model cannot learn a pick-and-place task due to the different modalities of pick and place trajectories. This issue is much more severe for reactive LfD approaches, such as dynamic systems, whose models further suffer from stability constraints. This work tries to solve this issue by presenting a task-parameterized energy function (TP-EF), which serves as a unified stability certificate across different tasks. The TP-EF can be learned on a mixture of tasks and generate stability certificates for different seen and unseen tasks. Building on the TP-EF, a task-parameterized dynamic system (TP-DS) is formulated to encode different task skills with strict stability guarantees and powerful reactive ability. As a result, the TP-DS serves as a unified skill model that can generate trajectories online for different seen and unseen tasks.

Förderung

  • Nanyang Technological University, Singapore
  • Ministry of Education, Singapore

Zitieren

APA

Jin, Z., Si, W., Ran, X., Yang, C., & Lv, C. (2026). Learning a Unified Dynamic System Model Across Diverse Robotic Demonstration Tasks. IEEE Transactions on Automation Science and Engineering, 23, 14290-14305. https://doi.org/10.1109/tase.2026.3718280

Vancouver

Jin Z, Si W, Ran X, Yang C, Lv C. Learning a Unified Dynamic System Model Across Diverse Robotic Demonstration Tasks. IEEE Transactions on Automation Science and Engineering. 2026;23:14290-14305. doi:10.1109/tase.2026.3718280

BibTeX
@article{Jin2026,
  author = {Jin, Zhehao and Si, Weiyong and Ran, Xu and Yang, Chenguang and Lv, Chen},
  title = {Learning a Unified Dynamic System Model Across Diverse Robotic Demonstration Tasks},
  journal = {IEEE Transactions on Automation Science and Engineering},
  year = {2026},
  volume = {23},
  pages = {14290--14305},
  doi = {10.1109/tase.2026.3718280}
}
Kostenlos starten →