Researchers at the Italian Institute of Technology used a convolutional neural network to reduce the amount of time the HyQ quadruped robot needs to plan its foot placement by several orders of magnitude.
Using the new technique, the robot can now make dynamic adaptations, enabling it to better withstand potentially destabilizing external forces.
The new controller permits HyQ to replan almost continuously, facilitating adjustments in real time, even when in the middle of a step.
The convolutional neural network was trained on terrain templates including gaps, bars, rocks, and other obstacles to interpret a three-dimensional map of the area before it, which is created by its onboard sensors.
The network is up to 200 times faster than traditional planning systems in terms of computing footstep selection.
From IEEE Spectrum
View Full Article
Abstracts Copyright © 2018 Information Inc., Bethesda, Maryland, USA
No entries found