This is the official repository for the implemntation of simulation experiments of CRIL: Continual Robot Imitation Learning via Generative Dynamics Model by Chongkai Gao, Haichuan Gao, Shangqi Guo, Tianren Zhang and Feng Chen.
Table of Contents
CRIL is a specialized deep generative replay algorithm designed for continual robot imitation learning that employs both a dynamics predictor and WGAN-GP for trajectory replay. The results of simulation and realworld experiments are as follows:
The replayed images of CRIL are as follows:
The simulation experiments of CRIL are based on MuJoCo and Meta-World benchmark, which need to be installed in advance. You can follow these instructions to install mujoco-py and meta-world.
Run the following code to collect rollouts for different tasks:
python3 data_collect.py
Run the following code to train the models:
python3 main.py
Note: you cannot run with only one click for various reasons. See the instructions in main.py.
We would like to thank Xin Su, Zhile Yang and Yizhou Jiang for various discussions on DGR theory and experiments of GANs. This work was supported in part by the National Natural Science Foundation of China under Grant 61671266 and Grant 61836004, in part by the Tsinghua-Guoqiang research program under Grant 2019GQG0006, and in part by Qualcomm Technologies, Inc.