Walk Calibration in a Four-legged Robot

5 Conclusions and future work

The use of walk calibration using machine learning proved to be feasible, as well as necessary, specially when changing the walking surface. The TeamChaos improved its maximum speed a 30% using simulated annealing. On the other hand, using multiple parameters sets allowed us to improve the speed response by fixing the discontinuities in the speed space. We obtained our best result using 2 sets of parameters. Moreover, the precise error measures enable the odometry system with more precise information about the motion uncertainity.

Calibrating the walk parameters for forward speed only is a good solution for the RoboCup domain. Using the infrastructure of our experiments, a study on a more complete calibration could be done, by calibrating simultaneously forward, lateral and rotational walking. This is not necessary in our domain and it would take much more time; still it is possible because of our instant speed measuring procedure. Finally, a calibration on curve lines could be considered, instead of separating calibration into three different walking types.

This research is funded by the project DPI2004-07993-C03-02 of the Spanish Government.