On Unsupervised Learning of Traversal Cost and Terrain Types Identification Using Self-organizing Maps

Learning-Based Methods of Perception and Navigation for Ground Vehicles in Unstructured Environments: A Review

“…In the past, several works have been proposed for the in-situ classification of terrain types, based exclusively on proprioceptive data while actually experiencing the traverse on the (potentially dangerous) terrain [ 88 , 89 ]. Among the works reported in the present survey, a recent and more frequently adopted approach consists of using proprioceptive data during training as well as for data labeling [ 34 , 35 , 66 ]. This approach has already proved to provide better classification and regression results compared to those methods based on exteroceptive data only, although only the latter are used at deployment time.…”

“…Faigl and Prágr [ 34 ] emphasised the need to define generalizable traversal costs and, with this idea, they considered instantaneous power consumption, maximum achievable forward velocity, and attitude stability. These measures can be used regardless of the specific type of platform.…”

“…Beyond LfD methods, an unsupervised method is adopted by Faigl and Prágr [ 34 ] for traversability cost maps regression. In particular self-organizing maps are proposed, to predict three traversal costs, i.e., instantaneous power consumption, maximum achievable forward velocity, and attitude stability, based on the morphology and the appearance of the terrain, acquired through an RGB-D camera.…”

“…In some cases, the manufacturers offer software development kits that even allow real-time environment mapping, with both geometry and visual appearance. This kind of environment sensing has already been employed in several works reported in this survey [34,51,75] and is much less costly than 3D laser scanners. However, depending on the sensing technique adopted, there are some application scenarios in which RGB-D cameras cannot work whereas LIDARs can, such as in the dark.…”

Energy-Based Legged Robots Terrain Traversability Modeling via Deep Inverse Reinforcement Learning

“…Traversability labels in exteroceptive-based methods ignore the robot's internal states while in motion across the terrain; hence, the degree of traversability is unknown, and the labels are usually binary (i.e., untraversable and traversable). In contrast, proprioceptive approaches can model terrain traversability as a continuous variable by defining a traversability cost using ground reaction score [11], vibration [12] or stability [13] from a proprioceptive sensor modality. However, these metrics are chosen based on a user's domain knowledge and may not fully capture the traversability experienced by the robot itself.…”