The objective of Loco3D is to build a methodology to execute a contact sequence, computed with an efficient motion planner, with a powerful humanoid robot inside a complex environment subject to dynamic changes. Planning and adapting complex locomotion patterns is a key problem that prevents from releasing humanoid robots and other legged manipulators in non-conventional factories, such as the assembly workshop of the Airbus A350 and A380 aircrafts in Toulouse, and beyond, the implementation of rehabilitation exoskeletons (e.g. for paraplegic) and the use service humanoid robots in offices, hospital and home environments.
In order to achieve this scientific objective, we divided the project in two stages, each one having its own scientific targets and outcomes. The first stage relies on mature scientific methods developed by our team, and serves as a basis for an original approach, in rupture with existing works, in order to tackle the locomotion problem in dynamic environments in a second stage. It also aims at bringing machine learning within the research topics of our team.
In the first stage, we will consider the nominal problem of locomotion in a complex static environment. The difficulty of this problem was emphasized by the recent DARPA Robotics Challenge and the spectacular falls of most of the biped robots involved. Based on some recent preliminary developments (on top of which the first real-time contact planner and the first real-time 3D locomotion pattern generator), we believe that our team is able to propose the premiere complete solution to this problem. The first part of the project will lead to the development of a first demonstrator that will be used as a springboard for the second phase of the project.
This first phase heavily relies on careful heuristics and domain-specific developments to handle locomotion in static environment. It might be possible to extend it to dynamic locomotion in uncertain environment, however asking again for years of development to find the proper heuristics and dedicated motion models able to handle these new hypotheses. Rather than doing the same work a second time in another context, we propose to study the issue of automatically generalizing the said heuristics in any motion context, by relying on the automatic construction of a robot motion experience.
Using the current know-how in robotics, it is relatively easy to automatically generate an efficient movement for any complex dynamics if the computation time is not considered. The question becomes much difficult when considering the hard time constraints imposed by the control frequency on a real hardware. A logical solution is to rely on pre-computation and storage of possible control values. However, the typical size of a naive storage is not acceptable. We propose in the second part of the project to transform the problem of motion generation in robotics into a problem of big data, by studying how to reduce the size of such a pre-computation dataset. The stake in data reduction is to exhibit a sparse structure in the dataset. The mathematical structure underlying the control function is explained at best by the Hamilton-Jacobi-Bellman equation. We propose to exhibit its sparsity by writing the reduction problem as an extended inverse optimal control problem.
The project is built on the applicative scenario of legged locomotion in complex dynamic environments. However, the general methodology is generic and would apply to the control of other complex dynamics. We will demonstrate the interest of using such an automatically-built robot experience by solving the problem of locomotion in dynamic environment, through the implementation of a complete demonstrator on the new humanoid robot of the laboratory. We will also demonstrate that it is generic by applying the same method to execute complex maneuvers with aerial manipulators in simulation.
Monsieur Nicolas Mansard (Centre National de la Recherche Scientifique/LAAS)
The author of this summary is the project coordinator, who is responsible for the content of this summary. The ANR declines any responsibility as for its contents.
CNRS/LAAS Centre National de la Recherche Scientifique/LAAS
Help of the ANR 253,716 euros
Beginning and duration of the scientific project: September 2016 - 36 Months