Inferring Distributions of Parameterized Controllers for Efficient Sampling-Based Locomotion of Underactuated Robots

Sampling-based motion planning algorithms provide a means to adapt the behaviors of autonomous robots to changing or unknown a priori environmental conditions. However, as the size of the space over which a sampling-based approach needs to search is increased (perhaps due to considering robots with many degree of freedom) the computational limits necessary for real-time operation are quickly exceeded. To address this issue, this paper presents a novel sampling-based approach to locomotion planning for highly-articulated robots wherein the parameters associated with a class of locomotive behaviors (e.g., inter-leg coordination, stride length, etc.) are inferred in real-time using a sample-efficient algorithm. More specifically, this work presents a data-based approach wherein offline-learned optimal behaviors, represented using central pattern generators (CPGs), are used to train a class of probabilistic graphical models (PGMs). The trained PGMs are then used to inform a sampling distribution of inferred walking gaits for legged hexapod robots. Simulated as well as hardware results are presented to demonstrate the successful application of the online inference algorithm.

I. Introduction

Autonomous systems currently suffer from an inability to safely control, or rather adapt, their behavior to achieve high-level goals in unstructured environments. In light of these limitations, one potential means to achieve the level of adaptation necessary for autonomous systems to successfully operate in unstructured environments is through sampling-based planning techniques (this potential is well documented in the motion planning literature). However, conventional sampling-based approaches tend to be computationally inefficient when sampling in high dimensional spaces, e.g., sampling the optimal parameters, relative to some high-level locomotive objective, for highly-articulated robots moving through uneven terrains. This work thus presents a new approach to sampling-based navigation planning for highly-articulated robots wherein a class of probabilistic graphical models is used to dramatically limit the effective size of the search space. More specifically, we show how to encode sets of features related to the kinematics, task objective, and environmental context in a PGM that is used within an online sampling-based inference algorithm to efficiently determine optimal motion parameters for underactuated robots moving through nontrivial terrains. Fig. 1.

Hexapod robot on which the efficient sampling-based planning framework developed in this work is demonstrated.