Learning to Drive Off Road on Smooth Terrains in Unstructured Environments Using an On-Board Camera and Sparse Aerial Images

Paper

Video Talk

Abstract

We present a method for learning to drive on smooth terrain in challenging off-road and unstructured outdoor environments. We extend an existing hybrid model-based and model-free reinforcement learning method to the outdoor driving domain. The prior method made efficient use of self-supervised collision detection to learn to quickly drive on free-space paths. Here, we use roughness, as measured using an IMU, as our learner’s reward. We provide both first-person and overhead aerial image inputs to our trained model and find that the fusion of both these complementary inputs makes for a policy with greater visual foresight for planning and is likewise more robust to different visual obstructions that might occur in either view. Our results show the ability to generalize smooth driving off road in the Canadian North, which contains plentiful vegetation, various types of rock, and sandy trails.

Citation

Download

@inproceedings{Manderson2020icra,
    title={Learning to Drive Off-Road on Smooth Terrains in Unstructured Environments Using an Onboard Camera and Sparse Aerial Images}, 
    author={Travis Manderson and Stefan Wapnick and Dave Meger and Gregory Dudek},
    year={2020},  
    booktitle = {Proceedings of the 2020 IEEE International Conference on Robotics and Automation}, 
    url = {https://arxiv.org/abs/2004.04697},
    year      = {2020}, 
    month     = {June}}