Pieter Abbeel, Gradescope, covariant.ai, and UC Berkley, USA
Title: High-performance model-based RL through learned domain randomization and meta-RL
Abstract: One of the main bottlenecks in robotic deployment is the human software engineering effort. Reinforcement learning is supposed to alleviate this effort, but in practice asymptotically optimal performance has only been achieved with model-free RL, which is extremely data inefficient. Model-based reinforcement learning approaches carry the promise of being data efficient. However, due to challenges in learning dynamics models that sufficiently match the real-world dynamics, they struggle to achieve the same asymptotic performance as model-free methods. I will discuss a new model-based RL approach (“model-based via meta-policy optimization (MB-MPO)”) that foregoes the strong reliance on accurate learned dynamics models. Using an ensemble of learned dynamic models, MB-MPO meta-learns a policy that can quickly adapt to any model in the ensemble with one policy gradient step. This steers the meta-policy towards internalizing consistent dynamics predictions among the ensemble while shifting the burden of behaving optimally w.r.t. the model discrepancies towards the adaptation step. Our experiments show that MB-MPO is more robust to model imperfections than previous model-based approaches. Finally, we demonstrate that our approach is able to match the asymptotic performance of model-free methods while requiring significantly less experience.
Aude Billard, EPFL, Switzerland
Title: Robust Control of Manipulation In Dynamic Environment
Abstract: I will present techniques developed in my group to enable robust and adaptive control of tasks that require to vary force and impedance upon contact. This enables robots to perform insertion and polishing tasks, and to peel vegetables, while adapting the control to the object moving during the manipulation. This also allows them to work collaboratively with humans.
Francesco Nori, DeepMind, UK and Istituto Italiano di Tecnologia (IIT), Italy
Title: Learning to endow robots with the ability to control physical collaboration through intentional interaction
Abstract: This talk focuses on the role of data collection and learning in the context of industrial applications. The talk will primary focus on the research conducted within the context of the European project AnDy. AnDy aims at measuring and modeling human whole-body dynamics to provide robots with an entirely new level of awareness about human intentions and ergonomy. AnDy has validated its progress in several realistic scenarios. In the first validation scenario, the robot is an industrial collaborative robot, which tailors its controllers to individual workers to improve ergonomy. In the second scenario, the robot is an assistive exoskeleton which optimizes human comfort by reducing physical stress. Within this context, AnDy is progressing along three technological and scientific breakthroughs. First, AnDy is innovating the way of measuring human whole-body motions by developing the wearable AnDySuit, which tracks motions and records forces. Second, AnDy learning models which combine ergonomic and cognitive predictive models (AnDyModel). Third, AnDyControl is an innovative technology for assisting humans through predictive physical control.
Zoe Doulgeri, Aristotle University of Thessaloniki, Greece
Title: Progressive Automation: A framework for learning by demonstration of repetitive and periodic tasks
Abstract: Robot learning from demonstration aims to reduce the programming time compared to conventional offline and online programming methods thus rendering collaborative robots viable for small and medium-sized enterprises where changes in the production line are frequent. The progressive automation framework allows a human to easily program a repetitive or periodic robotic motion task with kinesthetic guidance. The methodology actively assists the human while the robot learns to execute the task autonomously and allows seamless transition from guidance to autonomous operation. This is achieved by the synergetic action of Dynamic Movement Primitives, virtual fixtures and variable impedance control. I will present the techniques that were developed in my group to achieve synchronization between the DMP evolution and the human demonstration, variable impedance control methods that combine position and orientation errors, and virtual fixtures that are penetrable and can drastically assist the human in achieving progressive automation of a task within a few seconds.
Dongheui Lee, Technical University of Munich, Germany
Title: Robot Learning for Repetitive and Structured Tasks
Abstract: Conventional industrial robots execute exactly known repetitive tasks without uncertainties. One of challenges in industrial robots for industry 4.0 is intuitive interfaces for robot programming and adaptation to a systematically changing environment. In this talk, I will introduce our recent research on robot learning from demonstrations which could be applied to industrial tasks. Especially effective kinesthetic teaching method, learning of structured tasks, effective haptic exploration strategy, and conditional skills will be discussed.
Mathias Bürger, Bosch Center for Artificial Intelligence, Germany
Title: Learning Robotics – An Industrial View on Applications and Needs
Abstract: We will review some of the challenges for robotics we are face in flexible manufacturing and derive some research challenges for learning robotics. In particular, we will describe the research challenges and the research approaches we explore to increase the acceptance of Learning from Demonstration in industrial settings. We will also discuss further challenges for robotics in Industry 4.0 context where learning promises to have major impact.
Patrick van der Smagt, Volkswagen Group, Germany