Year of defence: 2019

Manuscript available here

Abstract

The locomotion problem in robotics is usually written as a large optimization problem. This problem has many undesirable properties, such as large dimensionality, non-linear and non-convex cos! functions, nonlinear constraints, and so on. These diverse features make direct resolution difficult. A classic approach is to simplify numerical optimization by solving reduced models, built heuristically. As a result, these models have a reduced area of validity and are difficult to extend. Our work explores an approach thal attempts to solve this problem of general locomotion, while fully exploiting itsstructure to obtain an effective and feasible solution. Our goal is to provide effective resolution through the proper use of the problem structure. For this, we have generalized the use of traditional models in locomotion in a formulation th at can solve exactly the initial problem. Our contribution consists of a method for the generation of dynamic movements forthe whole body of the humanoid, by the use of alternating direction method of multipliers. The complete problem is separated into (i) its centroidal part, which is highly constrained and unstable, and is solved by a multiple-shooting SQP, and (ii) its Lagrangian part, which is more stable but larger, and is solved by Differentiai Dynamic Programming (DDP). An efficient implementation of the DDP makes it possible to obtain performances allowing the implementation on the real robot, and a use of the feasibility const! raints ("proxy") makes it possible to en sure the feasibility of the reduced models.Our concepts have been validated empirically by calculating and applying different dynamic motions for the humanoid robot HRP-2, and to the robot Pyrene in simulation.