Model-based Optimization for Robotics

IEEE RAS Technical Committee

Background

Robots and in particular humanoid robots are extremely complicated dynamical systems for which the generation of behaviors is no easy task, since the number of parameters to tune for a behavior is very high. But the challenges waiting for today’s robots require them to automatically generate and control a wide range of behaviors in order to be more flexible and adaptive to changing environments. Optimization or optimal control offers an interesting way to generate behaviors automatically based on elementary principles (cost functions, constraints). Moore’s law as well as recent developments in optimization algorithms and in particular real-time optimization make a wider application of algorithmic optimization a realistic option even for real-time control in complex robotic applications in the near future.

There is a growing interest in optimization for behavior generation in robotics as recent publications and workshops show. However, it can be observed that the level of optimization techniques used often lacks far behind the current state of the art in the optimization community and that only very simple optimization problems are solved, which prevents the optimization potential of existing robot platforms to be fully exploited. Other papers show very interesting optimization solutions for humanoid robots even for complex tasks with impacts or varying constraints such as weight lifting or fast walking. At the same time, it is still a major challenge for optimal control to actually apply the solutions robustly to real robots and to handle modeling errors.

 

Goals

The aim of the IEEE RAS Technical Committee is to promote the usage of optimization in robotics, since the algorithms are at a very promising state but a joint research and communication effort is required to develop them further for the specific robotics needs, demonstrate their usefulness on real systems and make them available to the robotics community.

The goal is to address the following optimization-related key scientific issues:

  • Establish model-based optimization approaches for the automatic generation of robot behavior and actions in complex situations, in particular generate more natural looking humanoid motions
  • Develop model-based real-time optimization / model-predictive control techniques for the online control of fast robot motions
  • Showcase fast and robust optimal control actually implemented on robots in various situations

Show how state of the art optimal control techniques can be used in
model-based reinforcement learning.