Generating and controlling motions for complex robot systems such as humanoid robots is a challenging task. Model-based optimization, reinforcement learning and movement primitives represent three different bio-inspired approaches to tackle this issue. In this workshop, state of the art methods of all three fields are presented, and a special focus is put on how to best combine them to take the advantages of all approaches. Particular attention is paid to the application of whole-body motions wit changing contacts such as walking and balancing motions. Optimal control takes models of the robots at different levels of complexity as well as environmental constraints into account and can be used to exploit the physical limits of a robot. Optimal control can be performed in the offline and online modes (the latter often referred to as model-predictive control), but in all cases the model-reality mismatch has to be taken care of. Reinforcement learning can be performed without any model, but based on the real system, but for complex systems it may be very difficult to find feasible solutions that are feasible at all, so many iterations may fail. Movement primitives are very common in robotics, especially for transferring motions from humans to robots, with many fundamentally different variants of primitives around. The workshop will present latest research in all of these areas as well as new ideas for combinations and their applications to challenging robotics problems.
Topics:
Challenging types of whole-body robot motions that require sophisticated motion generation and control methods
State of the art methods for optimization / optimal control for multi-phase motion generation
State of the art methods for (nonlinear) model-predictive control / online optimization of motions
State of the art methods for reinforcement learning
Model-free vs. model-based learning
Different types of movement primitives (kinematic, dynamic, muscle synergies) and different types of segmentation approaches and retargeting
Different approaches to transfer primitives from humans to robots
Dynamic and kinematic models of humans and robots
Model reduction for optimization
Combining optimization & learning methods
Combining optimization & movement primitives
Shared open questions in both reinforcement learning and optimal control approaches
Considering multibody dynamics and constraints in movement primitives
Successful implementation of new (combined) methods to robots
12月13日
2016
会议日期
注册截止日期
留言