Efficient Humanoid Contact Planning using Learned Centroidal Dynamics Prediction

Abstract

Humanoid robots dynamically navigate an environment by interacting with it via contact wrenches exerted at intermittent contact poses. Therefore, it is important to consider dynamics when planning a contact sequence. Traditional contact planning approaches assume a quasi-static balance criterion to reduce the computational challenges of selecting a contact sequence over a rough terrain. This however limits the applicability of the approach when dynamic motions are required, such as when walking down a steep slope or crossing a wide gap. Recent methods overcome this limitation with the help of efficient mixed integer convex programming solvers capable of synthesizing dynamic contact sequences. Nevertheless, its exponential-time complexity limits its applicability to short time horizon contact sequences within small environments. In this paper, we go beyond current approaches by learning a prediction of the dynamic evolution of the robot centroidal momenta, which can then be used for quickly generating dynamically robust contact sequences for robots with arms and legs using a search-based contact planner. We demonstrate the efficiency and quality of the results of the proposed approach in a set of dynamically challenging scenarios.

A Vision-Based Two-Stage Framework for Inferring Physical Properties of the Terrain

“…For example, legged robots can use more favorable leg stiffness control to adapt to different terrain after obtaining the terrain stiffness information [1,2]. They can achieve better body balance to walk safely on the terrain with the friction information [3][4][5]. Wheeled robots can thus avoid excessive wheel slip and sinking.…”

Footstep Planning for Autonomous Walking Over Rough Terrain

“…One of the areas our approach is most lacking is a consideration of the CoM dynamics, as in [10]. While it is unclear if the learned-type dynamic cost prediction presented in that work is necessarily the best approach, it is certainly promising for producing plans that are dynamically variable to the environment structure.…”

“…More recently, a traversibility metric has been used to guide the expansion of ANA* planners to help accelerate convergence for multicontact planning [9]. Additionally, it has been shown that by including an estimated dynamics edge cost, the resulting plans can better adapt to the demands of different environments on the center of mass (CoM) dynamics [10]. A planar region segmentation of the environment has also been used to extract a desired 2D body path plan for the robot, and directly used to compute the desired footholds by assuming flat ground [8].…”

Robust Humanoid Contact Planning With Learned Zero- and One-Step Capturability Prediction

“…We assume there is a 0.4 second long swing phase followed by 0.6 second double support phase for each contact transition. We follow our previous work [7], given S(s), S(s ), r(s) and l(s), we use neural networks to predict dynamic feasibility of the contact transition, and determine r(s ) and l(s ).…”

“…Our goal is to maximize P success (T cp ), which can be achieved by minimizing K k=1 −log (P success (ε k )). Therefore, we define c cap as c cap (s, s ) = −log (P success (ε(s, s ))) (7) With this definition of c cap , we can find a path with maximum success rate by minimizing the total capturability cost of the path, which is done by the ANA* algorithm. In practice, we set w cap w s , d(s, s ) to let ANA* focus on maximizing P success (T cp ).…”