Generalization of Optimal Motion Trajectories for a Biped Walking Machine Based on Machine Learning

The TORO DLR-Biped is a two-legged rigid multibody system with 25 degrees of freedom. Recent work focused on the development of a motion planner to generate optimal trajectories for the walking task [1]. The motion planner is based on nonlinear optimization, to treat the complex nonlinear motion constraints, as well as to allow for the minimization of an appropriate cost function, like the actuation energy. Off-line computations result in globally optimized solutions.For these complex tasks, nonlinear optimization suffers from two flaws: the first is that the computation time is too long for any implementation in real-time; the second is that a suitable initial guess of the optimization parameters is necessary in order to avoid local non-convergence to a solution. To avoid these problems, a typical approach is to use a look-up table. Furthermore, to fulfil a general motion task, for example the humanoid robot step length, a generalization of the optimized solutions is required.In [2] the idea of a look-up table was extended with different machine learning methods, including fourth-order nearest neighbour, Support Vector Machines and Gaussian Processes. The result was a generalization of the look-up table, in form of a mapping between the task space (in our case, the walking task) and the optimization parameter space.In [2] the addressed task was the ball catching by means of a fixed-base manipulator. It is of interest here to extend this trajectory generation paradigm to the walking task for TORO. The final outcome should be the capacity to make use of globally optimal trajectories in a useful time. The method will also be implemented and tested in the lab on the real robot.

[1] Werner, A., Lampariello, R., Ott., C., "Optimization-based generation and experimental validation of optimal walking trajectories for biped robots", IEEE/RSJ International Conference on Intelligent Robots and Systems 2012 (IROS 12), Vilamoura, Algarve, Portugal, October 2012.

[2] R. Lampariello, D. Nguyen-Tuong, C. Castellini, G. Hirzinger and J. Peters,
"Motion planning for the energy-optimal robot catching in real-time", at the IEEE International Conference of Robotics and Automation 2011 (ICRA)