Virtual Sensors for Walking Robots

8.6 Conclusions

The development and testing of a ground-detection system for walking machines has been described. Joint-position error, joint speed and joint-position signals have been chosen as a feasible set of inputs for a virtual sensor to estimate the forces exerted by the robot’s foot. These inputs can be drawn from the joint-position sensors available in most robotic systems, so the implementation of this virtual sensor does not result in an extra hardware burden.

Data-driven virtual sensors have been found to be a suitable choice for developing a common sensorial system for several robotic platforms. The use of feed-forward neural networks to process the input information and generate a convenient output has been proved experimentally to be an adequate solution in a real walking robot. A calibration procedure has also been established with the aim of facilitating the tuning of these virtual sensors. The calibration procedure allows the detection system to be calibrated with a minimal experimental setup in a short period of time, and the required experiments are quite similar to the normal locomotion process. The results obtained in this first approach have been found to be satisfactory and are currently being used in the SILO4 walking machine for ground detection. The good quality of the force estimates obtained with this virtual sensor makes it a feasible alternative to the use of physical sensors. Because this method is easy to adapt to other machines, it is a good choice for simplifying the hardware of a servo-controlled system, or for providing it with low-cost sensor redundancy.