I'm Stéphane Caron, a locomotion software engineer at ANYbotics working with fantastic ANYmal quadrupeds. Before that, I was researcher in humanoid locomotion at CNRS in the IDH team in Montpellier, France, and visiting its sister laboratory JRL in Tsukuba, Japan. In both groups I've had the chance to work with excellent HRP humanoid robots.
How can we make legged robots walk better? At first, I thought their main limitation came from walking pattern generation, and studied that question, before coming to the conclusion that the technological bottleneck lied elsewhere. I then switched from Python to C++ and developed an open-source walking controller to make the HRP-4 humanoid climb stairs. In 2019, we used it in an industrial demonstrator at the Airbus Saint-Nazaire factory with the robot going through a pre-defined scenario: locate the staircase, climb it, go to a table, pick up a workpiece, apply it to the fuselage of an aircraft, etc.
Yet, HRP-4 and its fellow robots have the potential to walk in even more general scenarios. In this big picture, I focus on improving their balancing and decision making strategies. These are exciting questions to work on!... How can we make legged robots walk better?
As academics, we want our works to be reproduced not only by experts, who are already advanced in their various paths of knowledge, but also by newcomers eager to climb up their own paths. For them, published papers ripe with field idiosyncrasies (which come naturally from compression, in the information theoretic sense, to a fixed number of pages) are not an efficient tool. That's why we should distribute compilable source code of our full works. Note the emphasis on "compilable": non-compilable source code is weaker and less likely to be used, for good reasons. By giving access to the source, we make sure knowledge gaps can be crossed by anyone who takes the time to work it out. Papers cannot be fully detailed on every point, but compilable source code has to.
On a publishing variant of this stance, I support the model of overlay journals, also known as “reviewing entities” or “Peer Community in”, where peer reviewing happens on pre-print repositories like arXiv or HAL. We should put out pre-prints first, then iterate on them in the open, taking into account feedback from colleagues and reviewers. The record of these iterations is also valuable information for newcomers who want to cross knowledge gaps.