Het generaliseren van door mensen gedemonstreerde robotvaardigheden: gebruik van invariante modellen voor manipulatietaken met contact

In robotics, contact tasks involve manipulating objects while in contact with the environment or other objects, and they are prevalent across various fields, including industrial automation, domestic robotics, and healthcare. In this thesis, we focus on the manipulation of a single object attached to a robot end effector in contact with another object in the environment. These tasks encompass a wide range of applications, such as assembly, surface following, and interacting with articulated objects like doors or drawers. Programming such tasks is particularly challenging because it requires precise coordination and control of both the object's motion and its interaction forces with the environment. Learning from Demonstration (LfD) has the potential to simplify the programming of contact-rich tasks by making it more intuitive, allowing non-expert users to teach robots without requiring in-depth knowledge of robotics. However, despite these advantages, LfD still faces several challenges. Generalizing learned behaviors to new, unseen situations remains difficult, as robots need to adapt to varying conditions without explicit programming. Furthermore, data collection is expensive and data efficiency is a key issue, requiring the system to learn effectively from a small number of examples. Additionally, ensuring accurate calibration between multiple sensors is crucial for achieving reliable performance. These challenges highlight the need for robust methods to improve the practicality and reliability of LfD in real-world robotic applications. The objective of this thesis is to provide invariant models that are efficient to build, easy to fine tune, and applicable to real-world robot tasks in contact. These task models are derived from both motion and interaction force data that are collected during human demonstrations of the task and that are represented in a new form, making it easier to learn and generalize the task by a robot. The first contribution of this work is the introduction of invariant descriptors for force trajectories in contact tasks, building on the duality between motion and force. This provides a novel representation of force trajectories that complements existing descriptors for motion trajectories. The work also offers robust methods and software for calculating these descriptors, even in the presence of noisy measurements, using optimal control. Additionally, the approach is experimentally validated using recorded data from human demonstrations of two contact tasks, yielding task representations that are invariant to reference frame calibration and sensor placement. The second contribution is the development of an automatic method for deriving the optimal reference frame, or optimal task frame, directly from motion and force data recorded during task demonstrations. This method eliminates the need for expert-defined frames or predefined candidates by using screw theory to generate frame candidates and a probabilistic approach to determine the optimal one. It is robust to uncertainties in geometric features and works across various types of motion and force tasks, enabling invariant, accurate, and versatile task representation. The method has been validated through experiments showcasing its effectiveness in supporting learning-based controller design for contact tasks. The third contribution is the introduction of a fast and inexpensive closed-form method for estimating the force/torque sensor offsets, as well as the gravity parameters, i.e. mass and center of mass of the rigid object mounted on the force/torque sensor. This method leverages screw theory and relies solely on force and moment measurements, hence without requiring accurate motion data. The approach has been validated through both simulations and practical experiments, demonstrating its effectiveness in real-world applications, such as gravity compensation for force control of contact tasks. The fourth and final contribution is the derivation of relations between invariant descriptors and their corresponding moving frames, each derived from data expressed with respect to one of the two interacting rigid bodies. Although these relations were originally developed at the beginning of the PhD program and provided valuable insights, they are presented as the final chapter of this thesis. This placement allows for a deeper reflection on the derived relations and their implications, while also enabling the integration of data and results from the earlier chapters related to the first and second contributions, offering a more comprehensive understanding of the topic.