Ro2En: Robust Neural Environment Encoder for Domain Generalization of Fast Motion Planning

This paper discusses a new issue named domain generalization of fast motion planning in 3D environments, which benefits agility-required robot applications such as autonomous driving and uncrewed aerial vehicle obstacle avoidance flight. The existing work shows that conventional spatial search-based planning algorithms cannot meet the real-time requirement due to high time costs. The end-to-end neural network-based methods achieve an excellent balance between performance and planning speed in the seen environments, but are hard to transfer to new scenarios. To overcome this limitation, we propose a novel Robust Environment Encoder (Ro2En) approach to domain generalization of fast motion planning. Specifically, by demonstrating the reconstructed environment, we find that the previous environment encoder cannot encode the volume information properly, i.e., a volume collapse ensues, which leads to noisy environment modeling. Inspired by this observation, a dual-task auto-encoder is developed. It can not only reconstruct the point cloud of the obstacles, but also align their geometric centers. Experiment results showed that in the new scenarios, Ro2En outperformed previous state-of-the-art conventional and neural alternatives with a much smaller performance variation.