I'm a researcher interested in robot learning, particularly data-driven optimal control, reinforcement learning and visuomotor behavior policies. In the past, I made HRP-4 humanoid robots (1.5 m, 40 kg) climb stairs at the Airbus Saint-Nazaire factory, and worked with ANYmal quadrupeds at ANYbotics AG. Lately, I've been developing the open-source Upkie wheeled biped robots. You can check out the ideas I've worked on in my scientific publications and their corresponding code on GitHub. I also toot and post notes to this website.
Previously, I did research on humanoid locomotion at CNRS and was locomotion team lead at ANYbotics AG. And before that, I studied computer science at the École Normale Supérieure during my M.Sc. and robotics at the University of Tokyo during my Ph.D.
CoRL 2025 workshop on Open-Source Hardware in the Era of Robot Learning
Thanks to all who joined the CoRL 2025 workshop on Open-Source Hardware in the Era of Robot Learning!
The day brought together researchers and makers from both academia and the industry to discuss open-source robot design, integration with reinforcement learning frameworks, data collection and generation. The full recording of the workshop is available in the video archive.
Robot locomotion
One way to develop robot behaviors in the lab is to aim for repeatable demos, but at some point this goal steers us away from generalizing to the real world. For instance, the LIPM walking controller I worked on for the HRP-4 humanoid could repeatedly achieve a relatively complex task like stair climbing, but it wouldn't know how to react to unscripted scenarios like banging its head on a ceiling. That was due in part to all its behaviors being implemented by control systems, each coming with its own set of parameters to tune, stitched together by a finite state machine. The complexity of "keeping it all working together" increased significantly with each new behavior added to the system.
I recommend working with open-hardware robots to experiment faster than on commercial humanoids or quadrupeds. In my research, I prototype with Upkie wheeled bipeds whose stack is fully open source, from high-level Python down to the low-level firmware code that runs on moteus motor controllers. Upkies have only six actuated joints, yet they have the key features and core problems of legged robots like balancing and handling contacts. We have used them for instance when working on contact detection or obstacle-avoidance from vision.
Open access
I support open access and was fortunate to be able to publish most of my research as open-access manuscripts and open-source software. PDF files linked from this website are peer-reviewed revisions of the articles I've co-authored. On top of open access, I advocate for overlay journals as a way to challenge the oligopoly of publishers, who act as rent-seeking gatekeepers in the scientific ecosystem.