Learning Sim-to-real for Deformable Object Dynamics and Control
||Learning Sim-to-real for Deformable Object Dynamics and Control|
||SNIC Small Compute|
||Robert Gieselmann <firstname.lastname@example.org>|
||Kungliga Tekniska högskolan|
||2022-05-31 – 2022-12-01|
Deformable bodies such as cables, food and textiles can be found in many industrial and domestic environments. Advancing automation has increased the need for autonomous handling and manipulation of these objects. However, deformable bodies often exhibit complicated mechanics and high-dimensional configuration spaces, making classical modeling and control approaches impractical. Machine learning has shown promising results in robot control, such as learning manipulation policies directly from visual data. Promising results have also been obtained in learning dynamics, representations, and control policies for deformable objects. However, a major limitation of existing learning-based methods is the need for large amounts of high-quality data. Unlike real robot data, simulation data can be obtained on a large scale in a relatively short time. In this project, we investigate the extent to which simulated data can be used to accelerate the learning of dynamics models and controllers for deformable objects. In particular, we are interested in investigating sim-to-real adaptation and meta-learning techniques to improve the learning not only in simulation, but also in the real world.