Reduced models for legged locomotion: applications to motion planning, model predictive control and collision avoidance
The most challenging problems that we face when trying to make legged robots move are non-linear and high dimensional.
As a result the successes achieved by our community often rely on approximations of the original problem. Such models either introduce restrictions on the applicability of the proposed methods or lead to failures (most likely both).
However, approximated models can provide relevant initial guesses to non-linear methods, improve the performance of a machine learning framework by reducing a high-dimensional search space, and sometimes even result in robust methods that accurately resolve the initial problem on their own.
In this presentation, I will present new developments on results that I presented last year regarding footstep planning for legged robots. I will recall how a combinatorial problem of exponential complexity can be effectively approximated as a simple convex optimisation problem through model approximation. I will then elaborate on how the approximation can be used to systematically find a near-optimal solution to the original problem.
I will also present how similar ideas can be used to efficiently address other aspects of legged locomotion, namely model predictive control and the generation of collision-free motions.