Back to search

Aristotle University Of Thessaloniki (AUTH)

Enhancing the generalization and adaptation capabilities of dynamic movement primitives during robot learning from human demonstrations and execution in dynamic environments

Abstract

dc:description

Advancements in robotics have spurred interest in deploying robots alongside humans, necessitating efficient methods for transferring human skills to robots. Programming by Demonstration (PbD), where humans teach tasks to robots through demonstrations, offers an intuitive solution. For encoding human demonstrations and generalizing them to novel task conditions, the use of non-linear dynamical systems has been propounded, with Dynamic Movement Primitives (DMP) and Dynamical Systems with Gaussian Mixture Models (DS-GMM) being two of the most prevalent and widely employed models. However, existing approaches have limitations in adapting to dynamic environments and generalizing beyond demonstrated scenarios. This thesis addresses these limitations by proposing enhancements to learning human demonstrations with non-linear dynamical systems. Firstly, a novel parameterization for DS- GMM encoding Cartesian trajectories ensures global stability and reduces computational complexity. Additionally, leveraging task symmetries reduces model complexity and the number of required demonstrations. The focus then shifts to DMP, where a new mathematical formulation is derived, that is proven to be equivalent to the classical DMP formulation retaining all its favorable properties, while enabling to derive the model’s reference position, velocity, and acceleration in linear closed form with respect to the DMP’s learnable parameters. This enables training using only position measurements, online parameter alteration, and reversible execution. These advancements facilitate precise task execution and recovery from sensor faults. Furthermore, an optimization framework incorporates kinematic constraints and obstacle avoidance in real-time, enhancing online trajectory planning. Moreover, a novel spatial generalization framework overcomes limitations of traditional DMP generalization. It extends to dynamic via-points and external signals while preserving motion smoothness and shape fidelity. The applicability of DMP is also extended to complex tasks by employing deep Convolutional Neural Networks (CNNs) to infer DMP parameters from raw image inputs. This enables encoding multiple motion patterns and considering scene structure in task execution. By training the CNN to extract task-relevant features, DMP generalization incorporates complex scene information. In summary, this thesis contributes to advancing robot learning by enhancing DS-GMM and DMP capabilities. The proposed enhancements address limitations in generalization, adaptability to dynamic environments, and dealing with complex tasks. By combining novel mathematical formulations, optimization frameworks, and deep learning techniques, the thesis offers a comprehensive approach to improving robot learning from human demonstrations in dynamic environments, pushing the boundaries of what robots can accomplish autonomously.

Degree

thesis:*
Grantor dc:publisher
Aristotle University Of Thessaloniki (AUTH)
Year dc:date
2023

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Sidiropoulos, Antonios
  • Σιδηρόπουλος, Αντώνιος

Subjects

dc:subject × 18

Rights

Language dc:language
eng

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/56397

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Sidiropoulos, Antonios; Σιδηρόπουλος, Αντώνιος. Enhancing the generalization and adaptation capabilities of dynamic movement primitives during robot learning from human demonstrations and execution in dynamic environments. Aristotle University Of Thessaloniki (AUTH), 2023. http://hdl.handle.net/10442/hedi/56397