Enhancing Supervised Terrain Classification with Predictive Unsupervised Learning

Abstract

Abstract-This paper describes a method for classifying the traversability of terrain by combining unsupervised learning of color models that predict scene geometry with supervised learning of the relationship between geometric features and traversability. A neural network is trained offline on hand-labeled geometric features computed from stereo data. An online process learns the association between color and geometry, enabling the robot to assess the traversability of regions for which there is little range information by estimating the geometry from the color of the scene and passing this to the neural network. This online process is continuous and extremely rapid, which allows for quick adaptations to different lighting conditions and terrain changes. The sensitivity of the traversability judgment is further adjusted online by feedback from the robot's bumper. Terrain assessments from the color classifier are merged with pure geometric classifications in an occupancy grid by computing the intersection of the ray associated with a pixel with a ground plane computed from the stereo range data. We present results from DARPA-conducted tests that demonstrate its effectiveness in a variety of outdoor environments.

Learning slip behavior using automatic mechanical supervision

“…Note that in our training setup, the slip measurements come from some unknown nonlinear functions ( Figure 2) and could not be simply clustered into well discriminable classes, as previously done for characterizing terrains from mechanical vibration signatures [5], [8], or for learning terrain traversability in self-supervised learning [6], [11], [14], [17]. So, using these slip measurements as supervision is not a trivial extension of supervised learning.…”

“…Self-supervised learning uses one type of sensor to enhance the performance of another (e.g. learn range information from color features) and has been applied to extending the effective perception range [6], [11], [14], [17], [20], [22]. The above mentioned approaches use manual data labeling [11], [20], assume the sensor used as supervision can provide reliable labeling of terrain types [6], [14], or use heuristically defined traversability cost [22].…”

Identifying vegetation from laser data in structured outdoor environments

“…In the context of the LAGR-program (Learning Applied to Ground Robots), a number of methods were developed that exploit terrain knowledge in the surrounding of the robot to predict surface terrain in the far range. These near-to-far approaches use color information [18,19], 3D geometry information [20], or texture information [21,22]. Self-supervised learning was also used by Dahlkamp et al [23] in a vision-based road detection system.…”

Dimensionality Reduction Using Automatic Supervision for Vision-Based Terrain Learning

“…However, current methods are not automated enough and human supervision or some other heuristics are still needed to determine traversability [9], [16]. Recently, the concept of learning from the vehicle's sensors, referred to as learning from proprioception [16], or selfsupervised learning [4], [13], [19], has emerged.…”

“…Recently, the concept of learning from the vehicle's sensors, referred to as learning from proprioception [16], or selfsupervised learning [4], [13], [19], has emerged. This idea has proved to be particularly useful for extending the perception range [4], [9], [16], [19] which is crucial to increasing the speed and efficiency of the robot [4]. Self-supervised learning approaches require good separability in the space of sensor responses, so that a unique terrain class assignment for each example is obtained.…”