Visual Foothold Adaptations Using Machine Learning
Legged robots are becoming increasingly skilled and more often we see these machines performing tasks and experiments outside the confinements of a laboratory. Quadrupeds in particular are leading the charge to overcome the barriers of controlled environments. Nowadays, the number of fully untethered quadrupeds is rising, allowing them to encounter a wider range of situations.
Combining vision-less reactive behaviors and the use of visual information to prevent and plan for the future is pushing even further the capabilities of robots. We enhanced our Reactive Controller Framework (RCF) to adapt to the morphology of the terrain using a Convolutional Neural Network (CNN). Using this strategy, our robot HyQ is able to select proper footholds avoiding undesired situations such as stepping into gaps and leg collisions.
Selected publication:O. Villarreal, V. Barasuol, M. Camurri, L. Franceschi, M. Focchi, M. Pontil, D. G. Caldwell, C. Semini, Fast and Continuous Foothold Adaptation for Dynamic Locomotion through CNNs IEEE Robotics and Automation Letters, 1-1, 2019. View at publisher