モンテカルロ法で現実ギャップを埋める
A Monte Carlo Approach to Closing the Reality Gap
シミュレーションと実世界の観測から条件付き確率関数を構築し、状態空間カーネルでシミュレーションを現実に近づける手法を提案。改良したシミュレーションで学習した制御方策が実世界でも性能向上することを示した。
著者: Damian Lyons, James Finocchiaro, Michael Novitzky, Christopher Korpela
分類: cs.RO
原文アブストラクト
We propose a novel approach to the 'reality gap' problem, i.e., modifying a robot simulation so that its performance becomes more similar to observed real world phenomena. This problem arises whether the simulation is being used by human designers or in an automated policy development mechanism. We expect that the program/policy is developed using simulation, and subsequently deployed on a real system. We further assume that the program includes a monitor procedure with scalar output to determine when it is achieving its performance objectives. The proposed approach collects simulation and real world observations and builds conditional probability functions. These are used to generate paired roll-outs to identify points of divergence in behavior. These are used to generate {\it state-space kernels} that coerce the simulation into behaving more like observed reality. The method was evaluated using ROS/Gazebo for simulation and a heavily modified Traaxas platform in outdoor deployment. The results support not just that the kernel approach can force the simulation to behave more like reality, but that the modification is such that an improved control policy tested in the modified simulation also performs better in the real world.
