Training a terrain traversability classifier for a planetary rover through simulation

Abstract

A classifier training methodology is presented for Kapvik, a micro-rover prototype. A simulated light detection and ranging scan is divided into a grid, with each cell having a variety of characteristics (such as number of points, point variance and mean height) which act as inputs to classification algorithms. The training step avoids the need for time-consuming and error-prone manual classification through the use of a simulation that provides training inputs and target outputs. This simulation generates various terrains that could be encountered by a planetary rover, including untraversable ones, in a random fashion. A sensor model for a three-dimensional light detection and ranging is used with ray tracing to generate realistic noisy three-dimensional point clouds where all points that belong to untraversable terrain are labelled explicitly. A neural network classifier and its training algorithm are presented, and the results of its output as well as other popular classifiers show high accuracy on test data sets after training. The network is then tested on outdoor data to confirm it can accurately classify real-world light detection and ranging data. The results show the network is able to identify terrain correctly, falsely classifying just 4.74% of untraversable terrain.

Tutorial Review of Bio-Inspired Approaches to Robotic Manipulation for Space Debris Salvage

“…The forward model neural network was trained with both a backpropagation algorithm with momentum and an extended Kalman filter algorithm but only the latter could incorporate previous training data. The extended Kalman filter learning rule is given by [226]: This demonstrates the basic principle of forward models on a planetary rover pan-tilt camera mast with and without feedback in slewing the camera to follow visual targets. The adoption of a forward predictive model significantly reduces error excursions.…”

“…The forward model neural network was trained with both a backpropagation algorithm with momentum and an extended Kalman filter algorithm but only the latter could incorporate previous training data. The extended Kalman filter learning rule is given by [226]:…”

“…The forward model neural network was trained with both a backpropagation algorithm with momentum and an extended Kalman filter algorithm but only the latter could incorporate previous training data. The extended Kalman filter learning rule is given by [226]: over rugged terrain [225]. It compensated for rover/manipulator motion using a neural networktrained predictor of the pan-tilt dynamics emulating the VOR.…”

“…Unfortunately, these artificial muscle assemblies are cumbersome. Furthermore, biological viscoelastic properties are limited due to the low stiffness of muscles and tendons requiring internal models to compensate [226]-there are even more severe limitations in artificial muscles.…”

Supervised Learning of Natural-Terrain Traversability with Synthetic 3D Laser Scans

“…Moreover, synthetic depth data offers interesting opportunities for training traversability [19]. In this sense, virtual Lidar data generated with Matlab has been employed to build a neural network that classifies traversable terrain of planetary surfaces [20]. Similarly, in [21], a convolutional neural network has been trained to distinguish traversable patches from heightmap images obtained by the robotic simulator Gazebo [22].…”