Learning Food-arrangement Policies from Raw Images with Generative Adversarial Imitation Learning

In this paper, we tackle the problem of food- arrangement planning by an imitation learning approach from expert demonstrations. Specifically, we utilize a Generative Adversarial Imitation Learning framework, which allows an agent to learn near-optimal behaviors from a few expert demonstrations and self explorations without an explicit reward function. To evaluate our method, a food-arrangement simulator for the Japanese cuisine "Tempura" was developed with 3D-scanned tempura ingredients, and experiments were conducted for its performance evaluation. The experimental results demonstrate that our method can learn expert-like arrangement policies from bird-view raw images of plates without manually designing a reward function or requiring a massive number of expert demonstration data. Moreover, we confirmed that the learned polices are robust against arrangement errors and environmental changes compared to a baseline policy with supervised learning.