Speakers

Abstract: To be deployed in our everyday life, robots must display outstanding learning and adaptation capabilities allowing them to act, react, and continuously learn in unstructured dynamic environments. In addition, robots should display such capabilities in real time, which entails the ability to continuously learn from small numbers of demonstrations and/or interactions. In this context, the quality and efficiency of robot learning approaches may be improved via the introduction of inductive bias. In this talk, I will view inductive bias through the lens of geometry, which is ubiquitous in robotics. Specifically, I will discuss via three examples how geometry-based inductive bias can be introduced into robot learning from data structures, from physics, and from human knowledge. First, I will show that the performance of various algorithms may be improved by considering the intrinsic geometric characteristics of data. Second, I will discuss how the dynamic properties of humans and robots are straightforwardly accounted for by considering physics-based geometric configuration spaces. Finally, I will show that imposing an additional geometric structure to latent spaces allows us to learn low-dimensional representations of robotics taxonomies in continuous domains.

Title: Abstraction as Inductive Bias: Open-World 3D Scene Understanding without Open-World 3D Data

Title: Integrating Learning and Search for Task and Motion Planning Problems

Abstract: The recent advances in robot learning have been largely fueled by model-free deep reinforcement learning algorithms. These black-box methods utilize large datasets and deep networks to discover good behaviors. The existing knowledge of robotics and control is ignored and only the information contained within the data is leveraged. In this talk, we want to take a different approach and evaluate the combination of knowledge and data-driven learning. We show that this combination (1) enables sample-efficient learning on physical robots and (2) that generic knowledge from physics and control can be incorporated in deep network representations. The use of inductive biases for robot learning yields robots that learn dynamic tasks within minutes and robust control policies for under-actuated systems.

Title: From Structure to pre-trained representations: Inductive Biases for Robotics

Abstract: In this talk, I will cover some of our recent work exploring how to use graph networks to replicate impressive aspects of human physical problem solving, like our ability to create new physical tools. Using the relational inductive biases of graph networks, we can better model real world object dynamics and flexibly create tools like containers, funnels and even landscapes to solve new physical problems. This brings us closer to developing human level physical reasoning capabilities in machines.

Title: Skill-based Robot Learning

Title: Using Passive Internet Videos for Bootstrapping Robotics in the Wild

Title: Inductive Bias for Interaction with Humans