Many applications in robotics would benefit from robots being able to learn manipulation skills from only few demonstrations or trials. This contrasts with the ongoing trend in machine learning of constantly increasing the amount of data required to learn tasks. The main challenge of acquiring manipulation skills from limited training data is to find inductive biases and representations that can be used in a wide range of tasks, which requires us to advance on several fronts, including data structures and geometric structures. As example of data structures, I will discuss the use of tensor factorization techniques that can be used in global optimization problems to efficiently extract and compress information, while providing diverse human-guided learning capabilities (imitation and environment scaffolding). As examples of geometric structures, I will discuss the use of Riemannian geometry and geometric algebra in robotics, where prior knowledge about the physical world can be embedded within the representations of skills and associated learning algorithms.
Sylvain Calinon is a Senior Research Scientist at the Idiap Research Institute and a Lecturer at the Ecole Polytechnique Fédérale de Lausanne (EPFL). He heads the Robot Learning & Interaction group at Idiap, with expertise in human-robot collaboration, robot learning from demonstration and model-based optimization. The approaches developed in his group can be applied to a wide range of applications requiring manipulation skills, with robots that are either close to us (assistive and industrial robots), parts of us (prosthetics and exoskeletons), or far away from us (shared control and teleoperation).
This talk will take place as part of SCIoI member Svetlana Levit’s seminar “Selected Topics in Robot Learning,” which explores how advances in machine learning are helping robots operate in new environments, learn new behaviors, and adapt to changing conditions.
Image generated with DALL-E by Maria Ott.