|Organizers:||Ashis Gopal Banerjee, Evangelos Theodorou, Jan Peters|
|Location:||San Francisco, California, USA|
To equip humanoid robots and autonomous vehicles with enhanced real-time navigation and manipulation capabilities such as the ability to move over rough terrains, perform high-risk maneuvers, and grasp complex-shaped objects, robotics research has expanded from using machine learning for planning to more fine-grained control under uncertainty. The high dimensionality, hybrid nature of the underlying dynamics, partial observability, and lack of accurate models are just few of the challenges. Today's robots have highly nonlinear dynamics while the uncertainty is far from additive and Gaussian. These characteristics break any separation principles among stochastic estimation, feedback planning, optimal control and learning, and create the need for new methods that rely on all the frameworks. Such methods are being used to answer questions like: How can control theory be applied to improve long-horizon planning or state estimation performance? How can computation and execution of control policies in high dimensional state spaces be phrased as statistical learning or approximate inference problems?
This workshop will try to bring together researchers from both machine learning and controls communities to discuss state-of-the-art and identify promising research directions to provide greater adaptiveness and robustness to the next-generation robots.
The workshop is supported by the PASCAL2 Thematic Programme on Machine Learning for Autonomous Skill Acquisition in Robotics and the IEEE RAS Technical Committee on Robot Learning.
Ashis Gopal Banerjee, MIT
Evangelos Theodorou, University of Washington
Jan Peters, Max Planck Institute for Biological Cybernetics