I'm Stéphane, a robotics researcher in the Willow team at Inria Paris working on visual locomotion and manipulation. I have developed motion control software for HRP-4 humanoids, ANYmal quadrupeds, and now Upkie wheeled bipeds. You can read the algorithms in the corresponding publications or grab the code on GitHub. Occasionally, I toot and post technical notes to this website. Don't hesitate to send me an e-mail.
What's keeping walking machines from being useful? Perhaps it has to do with the way we design their behaviors, which is geared toward repeatable demos and mildly generalizes to the variability of the real world. For example, the walking controller I worked on for the HRP-4 humanoid achieved a relatively complex task like stair climbing, but it wouldn't know how to react to unscriped situations like banging the head on a ceiling (crouch) or falling forward after missing a step (extend arms). All the strategies it knew were implemented by control systems, where each behavior comes with its own parameters to tune, and code complexity increases with each new behavior.
I search for ways to go beyond this complexity ceiling. On that path, I started building Upkie, a wheeled biped that is fully open source and can be built at home. Upkie is a great testing partner to try all kinds of things quickly, with most of the code running in Python on a Raspberry Pi. That design choice turned out to be OK, especially for prototyping, as balancing and locomotion can be performed at remarkably low frequencies. This holds for the various balancers distributed with Upkie, whether they are based on PID control, model predictive control or reinforcement learning.
I'm a strong supporter of open access and was fortunate to be able to release most of my research as pre-prints and source code that I can keep maintaining. On top of open access, I see three ways in which we roboticists can work better as a community: overlay journals, post-prints, and raising our standards in terms of code distribution.