As for multi-legged wheel-legs, they were originally used as passive or minimally actuated devices to study walking metastability, stochasticity and dynamics on rigid ground [122, 123] and their ability to overcome obstacles using visco-elastic spokes [121, 124], shown in Fig. 2.14 (a). Their performance on rough terrain has been dealt with on several projects using DFKI’s Asguard rover. However, rather than modelling wheel-leg-soil interaction, these studies focus on higher level applications such as self-localisation using embodied data [125] and navigation and slip kinematics [126].
2.4
Machine Learning Solutions for Terrain Classification
Machine learning can be a powerful tool for terrain trafficability assessment, classifying unknown terrain into differentiated classes according to its geometric or physical charac- teristics. A brief introduction to the classification problem is given, presenting different approaches and algorithms. Previous research on the terrain classification problem for planetary exploration is presented, comparing different approaches for wheeled and legged robots. Finally, techniques for improving classification performance and robustness such as classifier fusion are reviewed.
Unlike terrain characterisation algorithms presented above, which estimate the values of terrain parameters through regression models, terrain classification algorithms categorise the terrain into distinct classes according to given criteria, e.g. the traversability or the type of terrain. For this purpose, terrain features are extracted from sensor data and passed on to the classifier algorithm, which decides to what class the terrain is more likely to belong.
All different approaches to classification follow a similar sequence, represented by the flow diagram in Fig. 2.15. First, relevant data is acquired from appropriately selected sensors. The continuous stream of data is then pre-processed by segmenting it into individual samples and transforming them into relevant features. These features are the input to the actual classification algorithm, which consists of two stages. The training phase consists of using readily classified feature datasets, i.e. data manually labelled according to the desired criterion, in order to learn the relationship between the sensed input features and the
2.4. Machine Learning Solutions for Terrain Classification
Figure 2.15: Generic flow diagram of the classification problem
class outputs. Once this relationship is modelled, the inference phase can apply it to new un-labelled sensed data in order to classify it in an automated fashion. Finally, post- processing of the classifier’s output can be done to minimise or reject miss-classifications or to produce an appropriate operational re-action, e.g. switch motion control modes or re-plan the path.
The main strength of the classification approach is that, thanks to machine learning tech- niques, complicated non-linear relationships can be implicitly modelled during the training phase, capturing them in a more accurate way than other modelling techniques such as the semi-empirical approach of Terramechanics presented in Section 2.3. On the other hand, as a Supervised Learning (SL) technique, the main weakness of classification algorithms is precisely their tight dependence on the training process.
Training datasets must be representative of the expected operating conditions of the clas- sifier, since its performance will significantly drop when trying to classify data that has not been appropriately trained to handle. In spite of this limitation, the learning capability of classifiers creates the opportunity of adapting to unknown conditions through new training phases.
This makes it a suitable solution for terrain modelling in planetary environments, where the uncertainty in soil characteristics and other factors, e.g. different gravity and atmospheric conditions, can produce unexpected behaviours and notable inaccuracies on a-priori models based on tests in Earth conditions. The dependency of classifiers on training datasets can be relaxed through Semi-Supervised Learning (SSL) techniques such as Transductive, Inductive or Active Learning [127].
2.4. Machine Learning Solutions for Terrain Classification
Regarding the classifier algorithms, five main categories can be distinguished: Nearest Neighbour Methods, e.g. k-Nearest Neighbours (kNN), Likelihood Methods based on Bayesian probability, e.g. Naive Bayes (NB), Discriminant Analysis methods (DA), e.g. Support Vector Machines (SVMs), Neural Networks (NNs) and Decision Trees (DT). Each of these methods offers a variety of algorithms and has its own advantages and disadvan- tages, being more or less suitable depending on the application. Some relevant factors affecting the choice of an appropriate algorithm are the dimensionality and heterogeneity of input data, the redundancy of the features and the complexity of the interactions between features. The performance of the algorithms can be evaluated in terms of classification accuracy, training time and classification time.
In the area of planetary exploration many researchers have proposed using classifiers as a way of providing semantic information about the terrain, which can be eventually related to terrain characteristics, e.g. trafficability. Input features for terrain classification can be divided in two groups: visual and tactile.
Visual features are obtained from cameras or ranging sensors which yield information about the texture or traversability [128, 129] of the terrain, as shown in Fig. 2.16 (a). They are easily portable between different rover platforms and can produce a useful classification of the terrain ahead of the rover. However, their outputs heavily depend on visibility conditions and are limited to geometrical and appearance information, which disregards many of the non-apparent properties of the soil and potential non-geometric hazards. Tactile features are obtained from IMUs, current transducers and force/torque sensors which return information about the vibrations and resistance forces produced by the terrain. Unlike the previous type of features they produce a classification based on direct measurements of the physical properties of the terrain as it reacts to the motion of the rover. The downside of these features is that they are dependent on the rover configuration and operation mode (load, size, speed, type of locomotion ...) and the classification algorithms need to be re-tuned for different rover platforms.
Various approaches have addressed tactile classification of terrain on wheeled rovers. Many of them rely on analyzing complex features of the vibrations produced by the terrain, such as Fast Fourier Transforms (FFT) used in [131] and Power Spectral Densities (PSD)
2.4. Machine Learning Solutions for Terrain Classification
used in [132], although in the latter study these features are also compared with simpler signal features. Principal Component Analysis (PCA) is often used to select the most discriminative features [133] and different classifiers such as SVMs [134] and NNs [135] have been applied for this task. The performance of various classification algorithms is compared in [136] and [137].
This performance can be improved using post-processing techniques like adaptive Bayesian Filters for temporal coherence [138], Unsupervised Learning (UL) techniques as done in [139] with an ad-hoc sensing probe to eliminate the need of extensive manually-labelled training datasets, the creation of novel classes of terrain using Gaussian Mixture Models (GMM) [140] and classifier fusion approaches to combine vibration- and vision-based classification [130, 141, 142], as depicted in Fig. 2.16 (b).
The tactile terrain classification problem has also been explored on legged rovers. The approaches found in the literature use force and torque sensors on the leg joints and foot tips [143, 144, 145], differences between goal and actual servo-motor [146] positions or on-board IMUs and back-EMF sensing [147]. Some applications also detect changes in gait bouncing to classify the terrain [148, 149]. In spite of the cited research, none of the approaches found in the literature uses multi-legged hybrid wheel-legs for terrain classifications. Moreover, the majority of the studies focuses on significantly different types of terrain, e.g. concrete, grass, tiles and gravel, rather than the narrower spectrum of dry sands addressed in this research.
(a) (b)
Figure 2.16: Terrain classification: (a) Vision-based terrain geometric hazard classification [128] and (b) vision-based classifier supervised by a vibration-based classifier [130]