A Comparison of Sampling Strategies for Parameter Estimation of a Robot Simulator

Abstract

Abstract. Methods for dealing with the problem of the "reality gap" in evolutionary robotics are described. The focus is on simulator tuning, in which simulator parameters are adjusted in order to more accurately model reality. We investigate sample selection, which is the method of choosing the robot controllers, evaluated in reality, that guide simulator tuning. Six strategies for sample selection are compared on a robot locomotion task. It is found that strategies that select samples that show high fitness in simulation greatly outperform those that do not. One such strategy, which selects the sample that is the expected fittest as well as the most informative (in the sense of producing the most disagreement between potential simulators), results in the creation of a nearly optimal simulator in the first iteration of the simulator tuning algorithm.

Fast damage recovery in robotics with the T-resilience algorithm

“…The most affected field is probably evolutionary robotics because of the emphasis on opening the search space as much as possible: behavior found within the simulation is often not anticipated by the designer of the simulator, therefore, it's not surprising that they are often wrongly simulated. Researchers in evolutionary robotics explored three main ideas to cross this 'reality gap': (1) automatically improving simulators (Bongard et al, 2006;Klaus et al, 2012;Pretorius et al, 2012), (2) trying to prevent optimized controllers from relying on the unreliable parts of the simulation (in particular, by adding noise) (Jakobi et al, 1995), and (3) modeling the difference between simulation and reality (Hartland and Bredeche, 2006;Koos et al, 2012).…”