拘束条件を考慮したバイラテラル制御に基づく模倣学習

抄録

In conventional methods of bilateral control-based imitation learning, restrictng a model’s output was infeasible once the model is trained. This led to safety issues, and hindered industrial application of bilateral control-based imitation learning. This study applied a variational autoencoder (VAE) to bilateral-control based imitation learning. A VAE is one of the encoder-decoder models with capability to present its latent variables as their Gaussian distributions. By sampling multiple sets of latent variables from one set of their distributions, generating multiple sets of commands from one set of input is made possible. In the experiments, this study verified constrained motion planning in bilateral control-based imitation learning using a VAE, by first generating command value using the average of derived latent variables, and then repeating attempts of sampling from their distributions, until reaching desired command value.