Trial-and-error learning for damage recovery with a quadruped robot
Auteur : Jean-Baptiste Mouret
|Téléphone||03 54 95 85 08|
Context and motivation
Our team works on new machine learning algorithms to allow robots to adapt to unexpected situations. For instance, we seek to make a hexapod robot that broke a leg re-learn to walk in a few minutes. We recently obtained promising results (see http://www.resibots.eu/videos.html and ).
We now wish to extend the adaptation possibilities by making our approach work on highly dynamic quadruped robot (https://www.youtube.com/watch?v=_YrWX9ez3jM). To do so, a new gait controller has to be designed and our learning algorithms have to be adapted.
The objectives of this internship are:
develop a fast simulator;
adapt the walking and learning algorithms for quadruped locomotion.
We aim at experimenting on the robots as much as possible but we have the support of 2 engineers. In case of blocking hardware issues, most of the work can be done in simulation.
use of ROS,
Cully, Antoine, Jeff Clune, Danesh Tarapore, and Jean-Baptiste Mouret. (2015) “Robots That Can Adapt like Animals”. Nature 521, no. 7553 (May 27, 2015): 503–7
Chatzilygeroudis K, Mouret JB. (2018) “Using parameterized black-box priors to scale up model-based policy search for robotics”. IEEE International Conference on Robotics and Automation (ICRA)