Abstract
Data-efficiency is crucial for autonomous robots to adapt to new tasks andenvironments. In this work we focus on robotics problems with a budget of only10-20 trials. This is a very challenging setting even for data-efficientapproaches like Bayesian optimization (BO), especially when optimizinghigher-dimensional controllers. Simulated trajectories can be used to constructinformed kernels for BO. However, previous work employed supervised ways ofextracting low-dimensional features for these. We propose a model andarchitecture for a sequential variational autoencoder that embeds the space ofsimulated trajectories into a lower-dimensional space of latent paths in anunsupervised way. We further compress the search space for BO by reducingexploration in parts of the state space that are undesirable, without requiringexplicit constraints on controller parameters. We validate our approach withhardware experiments on a Daisy hexapod robot and an ABB Yumi manipulator. Wealso present simulation experiments with further comparisons to severalbaselines on Daisy and two manipulators. Our experiments indicate the proposedtrajectory-based kernel with dynamic compression can offer ultra data-efficientoptimization.