WayFAST: Traversability Predictive Navigation for Field Robots

To deploy a field robot in wooded, obstacle prone, or poorly mapped areas, the user must plan the path around all possible obstacles. A recent paper on arXiv.org proposes a new modular approach, WayFAST (Waypoint Free Autonomous System for Traversability), to speed up field robotic path programming.

An example of an autonomous robot that is capable of crossing environmental obstacles.

An example of an autonomous robot that is capable of crossing environmental obstacles. Image credit: Sandia Labs/Randy Montoya, CC BY-NC-ND 2.0 via Flickr

The approach uses a modular architecture that combines traversability prediction using a convolutional neural network and a known kinodynamic model to autonomously navigate a variety of external environments. Learning traversable regions does not require any manual labels or heuristics to define it.

The method significantly reduces field robot path programming time and effort by allowing the user to choose only the target point where the robot needs to travel. Experiments in various challenging outdoor environments confirm that this method outperforms existing ones and leads to oscillation-free autonomous navigation.

We present a self-supervised approach to learning to predict traversable paths for wheeled mobile robots that require good traction to navigate. Our algorithm, called WayFAST (Waypoint Free Autonomous Systems for Traversability), uses RGB and depth data along with the navigation experience to generate autonomously traversable paths in an external unstructured environment. Our main motivation is that the traction for rolling robots can be estimated using kinodynamic models. Using the traction estimates provided by an online decreasing horizon estimator, we are able to train a traversability prediction neural network in a self-supervised manner, without the need for the estimates used by previous methods. We demonstrate the effectiveness of Wayfast through extensive field testing in a variety of environments, from sandy dry beaches to forest canopies and snow-covered grasslands. Our results clearly demonstrate that WayFAST can learn to avoid geometric obstacles as well as unremarkable terrain, such as snow, which would be difficult to avoid by sensors that only provide geometric data, such as LiDAR. Furthermore, we show that our training pipeline based on online traction estimations is more data-efficient than other heuristic-based methods.

Research Paper: Valverde Gasparino, M., “WayFAST: Traversability Predictive Navigation for Field Robots”, 2022. Link: https://arxiv.org/abs/2203.12071


Leave a Reply

Your email address will not be published.