Internships in Robotics

Send an email to jean-baptiste.mouret@inria.fr. More information is available from my web page.

Reactive controllers for hexapod robot

In our recent work (Cully et al., Nature, 2015 / see the video below), we demonstrated how a 6-legged robot can recover from an unforeseen damage conditions in less than 2 minutes. This novel machine learning algorithm opens many new possibilities to make robots more reliable and, overall, more adaptive.

Nevertheless, we used so far a very basic, open-loop controller (because our work was focused on the adaptation abilities). Our 6-legged robot is therefore unable to walk on rough terrains, whereas we would like our robot to behave as well as BigDog.

We recently added a 3D force sensor (OptoForce) on each foot of the robot, as well as a high-performance IMU. Now, we need to make a good use of them!

The first objective of this internship is to implement a reactive controller that makes use of these new sensor and thus improve the gait of the robot. The controller will be based on cartesian central pattern generator with sensor feedback [2].

The second objective is to test how using this controller affects the "Intelligent Trial and Error Algorithm" [1], which allows our robot to learn new gaits when needed.

The successful applicant will design new experiments and new algorithms to answer these questions. He/she will have access to the facilities of the lab (two 6-legged robots, Optitrack motion capture system, etc.) and he/she will be integrated in a highly-motivated team dedicated to leveraging trial-and-error learning to make robots that can adapt to anything (see: www.resibots.eu).

The ideal applicant loves robots. He/she has an appetite for machine learning algorithms and (modern) C++.

Video

References

  1. Cully, Antoine, Jeff Clune, Danesh Tarapore, and Jean-Baptiste Mouret. Robots That Can Adapt like Animals. Nature 521, no. 7553 (May 27, 2015): 503--507. doi:10.1038/nature14422.
  2. Barasuol, V., J. Buchli, C. Semini, M. Frigerio, E.R. De Pieri, and D.G. Caldwell. Reactive Controller Framework for Quadrupedal Locomotion on Challenging Terrain. In 2013 IEEE International Conference on Robotics and Automation (ICRA), 2554--2561, 2013. doi:10.1109/ICRA.2013.6630926.

Recovering from a sequence of damages with fast trial-and-error learning (Robotics / machine learning)

In our recent work (Cully et al., Nature, 2015 / see the video below), we demonstrated how a 6-legged robot can recover from an unforeseen damage conditions in less than 2 minutes. This novel machine learning algorithm opens many new possibilities to make robots more reliable and, overall, more adaptive.

But, what would happen if the robot is damaged again? Should the robot forget everything it has learned when trying to cope with the first damage condition? If not, what should it keep? And, what if the adaptation was not an actual damage but a perturbation from the enviroment, for instance a different terrain? How could the robot recognize that "the situations looks like something that it is has seen before" if it goes back to the initial terrain?

The successful applicant will design new experiments and new algorithms to answer these questions. He/she will have access to the facilities of the lab (two 6-legged robots, Optitrack motion capture system, etc.) and he/she will be integrated in a highly-motivated team dedicated to leveraging trial-and-error learning to make robots that can adapt to anything (see: www.resibots.eu).

The ideal applicant loves robots. He/she has an appetite for machine learning algorithms and (modern) C++.

Video

References

  1. Cully, Antoine, Jeff Clune, Danesh Tarapore, and Jean-Baptiste Mouret. Robots That Can Adapt like Animals. Nature 521, no. 7553 (May 27, 2015): 503--507. doi:10.1038/nature14422.

Probabilistic Motion Primitives Without Demonstrations

Probabilistic Movement Primitive (ProMP) is a recent concept to learn trajectories in robotics [1]. Starting with a few demonstrations from a teacher, a ProMP builds a probabilistic distribution of the demonstrations. This representation allows other algorithms to infer the best trajectory depending on the context (e.g. a movement by an operator) [2].

Instead of starting with demonstrations, we would like to generate the distribution of potential trajectories with an algorithm; we think the MAP-Elites algorithm [3][4], a novel evolutionary algorithm that we published last year, could be the ideal algorithm to do so.

The objective of the internship is to mix Probabilistic Motion Primitives with MAP-Elites. Most experiments will be performed on our Kinova robotic arm or with the iCub humanoid robot.

The successful applicant will design new experiments and new algorithms to answer these questions. He/she will have access to the facilities of the lab (two 6-legged robots, Optitrack motion capture system, etc.) and he/she will be integrated in a highly-motivated team dedicated to leveraging trial-and-error learning to make robots that can adapt to anything (see: www.resibots.eu).

The ideal applicant loves robots. He/she has an appetite for machine learning algorithms and (modern) C++.

References

  1. Paraschos, Alexandros, Christian Daniel, Jan Peters, and Gerhard Neumann. "Probabilistic movement primitives." In Advances in Neural Information Processing Systems, pp. 2616--2624. 2013.
  2. Maeda, Guilherme, Marco Ewerton, Rudolf Lioutikov, Heni Ben Amor, Jan Peters, and Gerhard Neumann. "Learning interaction for collaborative tasks with probabilistic movement primitives." In Humanoid Robots (Humanoids), 2014 14th IEEE-RAS International Conference on, pp. 527--534. IEEE, 2014.
  3. Mouret, Jean-Baptiste, and Jeff Clune. "Illuminating search spaces by mapping elites." arXiv preprint arXiv:1504.04909 (2015).
  4. Cully, Antoine, Jeff Clune, Danesh Tarapore, and Jean-Baptiste Mouret. Robots That Can Adapt like Animals. Nature 521, no. 7553 (May 27, 2015): 503--507. doi:10.1038/nature14422.

Learning to crawl with a damaged iCub robot

The iCub robot is a child-like humanoid robot (see http://www.icub.org) with 53 degrees of freedom, a sensitive skin, and many sensors. A previous project made iCub crawl on the ground [1]. The goal of this master internship is to use the Intelligent Trial and Error algorithm [2] to learn a new gait when the iCub is broken (e.g. a broken cable in the elbow).

The successful applicant will design new experiments and new algorithms to answer these questions. He/she will have access to the facilities of the lab (two 6-legged robots, Optitrack motion capture system, etc.) and he/she will be integrated in a highly-motivated team dedicated to leveraging trial-and-error learning to make robots that can adapt to anything (see: www.resibots.eu).

The ideal applicant loves robots. He/she has an appetite for machine learning algorithms and (modern) C++.

References

  1. Degallier, S., Righetti, L., Natale, L., Nori, F., Metta, G., & Ijspeert, A. (2008, October). A modular bio-inspired architecture for movement generation for the infant-like robot iCub. In Biomedical Robotics and Biomechatronics, 2008. BioRob 2008. 2nd IEEE RAS & EMBS International Conference on (pp. 795-800). IEEE.
  2. Cully, Antoine, Jeff Clune, Danesh Tarapore, and Jean-Baptiste Mouret. Robots That Can Adapt like Animals. Nature 521, no. 7553 (May 27, 2015): 503--507.