Recurrent and convolutional neural networks for deep terrain classification by autonomous robots

The future challenge for field robots is to increase the level of autonomy towards long distance (>1 km) and duration (>1h) applications. One of the key technologies is the ability to accurately estimate the properties of the traversed terrain to optimize onboard control strategies and energy efficient path-planning, ensuring safety and avoiding possible immobilization conditions that would lead to mission failure. Two main hypotheses are put forward in this research. The first hypothesis is that terrain can be effectively detected by relying exclusively on the measurement of quantities that pertain to the robot-ground interaction, i.e., on proprioceptive signals. Therefore, no visual or depth information is required. Then, artificial deep neural networks can provide an accurate and robust solution to the classification problem of different terrain types. Under these hypotheses, sensory signals are classified as time series directly by a Recurrent Neural Network or by a Convolutional Neural Network in the form of higher-level features or spectrograms resulting from additional processing. In both cases, results obtained from real experiments show comparable or better performance when contrasted with standard Support Vector Machine with the additional advantage of not requiring an a priori definition of the feature space.     

A piecewise bearing capacity method of unstructured terrain considering characteristics of soil mechanic and wheel geometry

Bearing capacity of the unstructured terrain considering the effects of the wheel geometry and soil mechanic properties is analyzed in this paper. Two-dimensional pressure-sinkage simulations are conducted to evaluate the degrees of similarity between the flat plate and wheel in terms of their ultimate bearing characteristics of Terzaghi theory. The results show that these degrees of similarity are mainly reflected in the soil in-depth direction and the corresponding failure behaviors. Based on the approximation of the ultimate bearing capacity between the wheel and flat plate, a piecewise bearing capacity evaluation method with the effects of the soil mechanic properties and three-dimensional wheel geometry is proposed. The pressure-sinkage values of the proposed model show a satisfactory agreement with the experimental ones. The proposed model performs better than the semi-empirical models, as it considers more soil bearing features and needs less fitting parameters to assess the unstructured terrain.     

Data processing methodology in the characterization of the mechanical properties of terrain

The methodology for processing load-sinkage and shear test data obtained using the bevameter is examined. A weigthed least squares technique is proposed for deriving terrain values from experimental data. To streamline data processing time, a portable automatic data processing unit for the bevameter has been developed. The basic features and functions of the unit are described. The automatic data processing unit, together with associated bevameter, has undergone extensive field trials. They have been employed to measure the properties of various types of natural terrain. The results indicate that this new system can provide a basis for the development of a reliable, convenient and standardized technique for identifying terrain properties in relation to off-road mobility.     

Exploring cold regions autonomous operations

The US Army must update its vehicle fleet to be better equipped for potential future military conflicts in northern climates (US Army, 2017). This process involves considering manned, optionally manned, and unmanned vehicles as viable options in the future. Optionally manned and unmanned vehicles in the armed forces have substantial benefits because they can operate without direct driver input or are able to perform missions deemed too dangerous for troops. Optionally manned vehicles allow the driver to shift some, or all, focus away from the task of driving the vehicle. In some cases, these autonomous vehicles may perform better than a human driver by rapidly sensing and reacting to terrain changes. Onboard sensing and decision making are equally applicable to both fully autonomous and teleoperated vehicles. This work will focus on the terrain sensing, waypoint navigation, and teleoperation potential of an optionally manned or unmanned vehicle. Results from a vehicle demonstration on two different terrain conditions will provide the basis for additional terrain sensing and autonomous vehicle development work in the coming year.     

Cross-country mobility on various snow conditions for validation of a virtual terrain

Realistic simulation of on- and off-road vehicle performance in all weather conditions is needed by the U.S. Army for virtual training of personnel on existing vehicles, and for new vehicle design. The virtual test site is a computer simulation representing an actual terrain defined as having spatially distributed terramechanics properties and terrain interaction with vehicles. We developed a virtual test site for Ethan Allen Firing Range (EAFR) in northern Vermont. The virtual test site for EAFR is composed of terramechanics properties including spatially distributed snow depth and density, soil type, drainage class, slope, and vegetation type. Snow depth and density were spatially distributed with regard to elevation, slope, and aspect using a surface energy balance approach. This paper evaluates whether the terramechanics representation of a virtual test site is improved by adding spatially distributed snow and soil properties, rather than using uniform properties. The evaluation was accomplished by conducting a cross-country vehicle performance analysis using the North Atlantic Treaty Organization (NATO) Reference Mobility Model (NRMM) to validate the new algorithms for realistic spatial distribution of snow properties. The results showed that the percentage of No-Go areas for uniform snow is lower than the distributed snow by 4% for the CIV (CRREL Instrumented Vehicle), 8% for the HMMWV (High Mobility Multipurpose Wheeled Vehicle), and 5% for the Stryker vehicle. For both light vehicles, approximately 12% of the No-Go areas are classified as such because of slopes 29%. These results imply that spatial distribution of snow properties provides realistic vehicle response as opposed to having the snow properties distributed uniformly throughout the entire terrain. This represents an improvement over previous versions of the terramechanics properties.     

Calculating fractal parameters from low-resolution terrain profiles

Driver comfort on rough terrain is an important factor in the off-road performance of wheeled and tracked ground vehicles. The roughness of a terrain has typically been quantified by the U.S. Army as the root-mean-square elevation deviation (RMS) of the terrain profile. Although RMS is an important input into many mobility calculations, it is not scale invariant, making it difficult to estimate RMS from low resolution terrain profiles. Fractal parameters are another measure of roughness that are scale invariant, making them a convenient proxy for RMS. While previous work found an empirical relationship between fractal dimension and RMS, this work will show that, by including the cutoff length, an analytic relationship between fractal properties and RMS can be employed. The relationship has no free parameters and agrees very well with experimental data - thus providing a powerful predictive tool for future analyses and a reliable way to calculate surface roughness from low-resolution terrain data in a way that is scale invariant. In addition, we show that this method applies to both man-made ride courses and natural terrain profiles.     

Modeling and Simulation of a Three-Wheeled Mobile Robot on Uneven Terrains with Two-Degree-of-Freedom Suspension Mechanisms

A wheeled mobile robot (WMR) will move on an uneven terrain without slip if its torus-shaped wheels tilt in a lateral direction. An independent two degree-of-freedom (DOF) suspension is required to maintain contact with uneven terrain and for lateral tilting. This article deals with the modeling and simulation of a three-wheeled mobile robot with torus-shaped wheels and four novel two-DOF suspension mechanism concepts. Simulations are performed on an uneven terrain for three representative paths—a straight line, a circular, and an ‘S’-shaped path. Simulations show that a novel concept using double four-bar mechanism performs better than the other three concepts.     

Comparisons on engine power requirements of 6WD and 4WD prime movers for the oil palm plantation in Malaysia

An attempt was made to investigate the possibility of designing and developing a multipurpose, all terrain, 6WD prime mover for the oil palm plantation in Malaysia. Comparisons were made on the engine power requirements of the 6WD prime mover over the 4WD prime mover on four different soil classifications having terrain slopes ranging from 0° to 30° based on the traction equations by ASABE [1]. Generally, the 6WD prime mover showed in the range of 5.27-45.81% reduction in the engine power requirements than the 4WD prime mover having equal size and weight configurations when traversing over the four different soil classifications. Greater percentage reductions in engine power with the 6WD over the 4WD prime mover were found as the terrain changes from concrete to soft or sandy soil classification with prominent percentage reductions at higher terrain slopes. The proposed 6WD prime mover has a single chassis with skid steer drive wheels, oscillating drive axles, and low inflation pressure tires. The 63 kW water-cooled, diesel engine was sufficient to run an hydrostatic main pump at a working pressure of 220 bar and a flow rate of 91 L/min under two drive speeds (i.e. high and low) and two drive modes (i.e. forward and reverse). Proper prime mover wheelbase and proper ground clearance height were employed to give better stability and manoeuvrability for the typical oil palm plantation terrain in Malaysia. Mounting provisions for the seedling transplanting, fertilizer applicator and the in-field fresh fruit bunches (FFB) collection-transportation were made on the prime mover.     

Practical method of improving the turnability of terrain vehicles

To improve the trafficability and the turnability of the terrain vehicles, it was already pointed out in a previous report that the control of the ground contact area of the running gear such as tracks and wheels could be highly recommended. The contact area of the running gear must be as large as possible to obtain more traction, however, less contact area would be better to obtain easier turning and steering. In this paper the principle of improving the turnability for the terrain vehicle was theoretically discussed. One of the examples of the practical application of the theory developed here was proposed and applied to the terrain vehicle equipped with eight powered wheels, which was constructed as a test vehicle for this study.     

Modeling of terrain impact caused by tracked vehicles

Analytical models that can predict the terrain impact caused by tracked vehicles on a horizontal plane were developed and tested. The models included a disturbed width model and an impact severity model. Inputs of the terrain impact models included vehicle static properties, vehicle dynamic properties, and terrain properties. The tested vehicles included an M1A1 tank, an M577 Armored Personal Carrier (APC), and an M548 cargo carrier. The models were verified with field tests conducted in Yakima Training Center in Yakima, WA, Fort Riley, KS, and Camp Atterbury, Indiana. The average percentage errors of the disturbed width model for the M1A1, M577, and the M548 were 10.0%, 27.3%, and 8.5%, respectively. The average percentage errors of the impact severity model of the M1A1 and M577 were 25.0% and 21.4%, respectively.     

Statistical modeling and comparison with experimental data of tire–soil interaction for combined longitudinal and lateral slip

The interaction of a tire with a soft terrain has multiple sources of uncertainties such as the mechanical properties of the terrain, and the interfacial properties between the tire and the terrain. These uncertainties are best characterized using statistical methods such as the development of stochastic models of tire–soil interaction. The quality of the models can be assessed via statistical validation measures or metrics. Although validation of stochastic tire–soil interaction models has recently been reported with good results, it involves longitudinal slip only without considering lateral slip which can occur simultaneously with longitudinal motion. This paper presents results of the validation of a simple stochastic tire–soil interaction model for the more complicated case of combined slip. The statistical methods used for validation include the development of a Gaussian process metamodel, the calibration of model parameters using the approach of the maximum likelihood estimate in conjunction with new test data. The validation of the calibrated model, when compared with test data, is obtained using four validation metrics with good results.     

Terrain trafficability analysis and soil mechanical property identification for planetary rovers: A survey

The advances in the field of robotics enabled successful exploration of the Moon and Mars. Over the years, rover missions have demonstrated deployment of various scientific payloads for robotic field geology on these extra-terrestrial bodies. The success of these missions clearly emphasises the need to further advance rover technology in order to maximise scientific return. The success of future robotic surface exploration missions will depend on two key factors – autonomy and mobility on soft sandy and unstructured terrains. The main contribution of this paper is that it brings together vital information pertaining to various terrain characterisation techniques into a single article. Special care is taken in structuring the paper so that all the relevant terrain characterisation methods that have been used in past planetary exploration missions and those under consideration for future space exploration missions are covered. This paper will not only lists advantages and disadvantages of various terrain characterisation techniques but also presents the methodology for evaluating and comparing terrain characterisation techniques and provides a trade-off study of existing and potential approaches that could improve the mobility of future planetary exploration rovers. This survey shows that further advances in currently deployed technology are required in order to develop intelligent, on-board sensing systems which will detect and identify near surface and sub-surface terrain properties to enhance the mobility of rovers.     

Reduced testing and modelling of the bearing capacity of rooted soil for wheeled forestry machines

The next generation of forestry machines must be developed to be gentler to soil and to the root mat than present machines, especially in thinning operations. The bearing capacity of the soil is a key property for determining the terrain trafficability and machine mobility. This asks for better and more general terramechanics models that can be used to predict the interaction between different machine concepts and real and complex forest soil.This paper presents results from terramechanics experiments of rooted soil with a new and small-scale testing device. The force–deflection results are analyzed and compared with analytical root reinforcement models found in literature. The presented study indicates that rooted soil properties obtained with the new laboratory test device can be used to create an augmented soil model that can be used to predict the bearing capacity of rooted soil and also to be used in dynamic machine–soil interaction simulations.     

Foot force distribution of walking machine with consideration of terrain properties

Determination of foot force distribution during walking is important to the simulation and control of the vehicle. This problem was often considered as an indeterminate problem and several optimization methods were proposed. The indeterminancy, which was due to the assumption of rigid bodies, can, however, be removed by incorporating the compliance equations into the equations of equilibrium. Based on such a compliant model, a s stiffness matrix method was developed to determine the foot force distribution. However, due to the complexity of the problem, the compliance of terrain in the stiffness matrix method was considered either negligible or as a linear spring model. In this paper, two realistic terrain models are incorporated into the stiffness matrix method to study the effect of terrain properties on foot force distribution during walking. These two terrain models are the three-parameter solid model and the four-parameter Burgers model. The former is a model of clay terrain while the latter is a model of a paddy field. These models are extended to three dimensions and then combined with the leg compliances to form the stiffness matrix of the system. The simulation results show that the terrain compliance has significant effect on foot force distribution. For example, it is observed that this compliance helps to distribute the foot forces evenly and to minimize the frictional angles.     

Terrain classification using intelligent tire

A wheeled ground robot was designed and built for better understanding of the challenges involved in utilization of accelerometer-based intelligent tires for mobility improvements. Since robot traction forces depend on the surface type and the friction associated with the tire-road interaction, the measured acceleration signals were used for terrain classification and surface characterization. To accomplish this, the robot was instrumented with appropriate sensors (a tri-axial accelerometer attached to the tire innerliner, a single axis accelerometer attached to the robot chassis and wheel speed sensors) and a data acquisition system. Wheel slip was measured accurately using encoders attached to driven and non-driven wheels. A fuzzy logic algorithm was developed and used for terrain classification. This algorithm uses the power of the acceleration signal and wheel slip ratio as inputs and classifies all different surfaces into four main categories; asphalt, concrete, grass, and sand. The performance of the algorithm was evaluated using experimental data and good agreements were observed between the surface types and estimated ones.     

Design and characterization of GRC-1: A soil for lunar terramechanics testing in Earth-ambient conditions

Earth experiments must be carried out on terrain that deforms similarly to the lunar terrain to assess the tractive performances of lunar vehicles. Most notably, terrain compaction and shear response underneath the lunar vehicle wheels must represent that of the Moon. This paper discusses the development of a new lunar soil simulant, Glenn Research Center lunar soil simulant #1 (GRC-1), which meets this need. A semi-empirical design approach was followed in which the soil was created by mixing readily available manufactured sands to a particle size distribution similar to the coarse fraction of lunar soil. By varying terrain density, a broad range of in situ cone penetration measurements collected by the Apollo mission astronauts can be replicated. An extensive set of characterization data is provided in this article to facilitate the use of this material. For reference, the index and geotechnical properties of GRC-1 are compared to the lunar soil and existing lunar soil simulants.     

Prediction of impacts of wheeled vehicles on terrain

Traffic of off-road vehicles can disturb soil, decrease vegetation development, and increase soil erosion. Terrain impacts caused by wheeled off-road vehicles were studied in this paper. Models were developed to predict terrain impacts caused by wheeled vehicles in terms of disturbed width and impact severity. Disturbed width and impact severity are not only controlled by vehicle types and vehicle dimensions, but also influenced by soil conditions and vehicle dynamic properties (turning radius, velocity). Field tests of an eight-wheeled vehicle and a four-wheeled vehicle were conducted to test these models. Field data of terrain–vehicle interactions in different vehicle dynamic conditions were collected. Vehicle dynamic properties were derived from a global position system (GPS) based tracking system. The average prediction percentage error of the theoretical disturbed width model is less than 20%. The average absolute error between the predicted impact severity and the measured value is less than an impact severity value of 12%. These models can be used to predict terrain impacts caused by off-road wheeled vehicles.     

Numerical analysis to predict turning characteristics of rigid suspension tracked vehicle

Numerical analysis was developed to calculate the steering properties of a rigid suspension tracked vehicle turning on soft terrain. The developed numerical analysis is based on a method to solve a set of non-linear equations. To verify the numerical analysis, an experiment on a model-tracked vehicle turning with a steering ratio of 1.6 on a loose sandy terrain was carried out. The comparison between measured and calculated values shows that the numerical analysis can predict sinkage, slip ratios and turning radius within an error amount of 15%.     

A comparative study of destructive effects resulting from road profile acting on off-road towed vehicles

The more extreme conditions the vehicle is exposed to, the sooner it wears out and deteriorates. In order to determine the forces affecting the lifespan of vehicles we need to know the environmental conditions eliciting these forces.This research aims at elaborating and testing a method which can help to conduct a comparative analysis of forces acting on towed vehicles used in different terrain conditions. Excitation forces acting on a vehicle being towed across terrain cause vibrations which lead to wear and structural deterioration. The rate of deterioration depends on the activating forces resulting from the road profile geometry and the dynamic properties of the vehicle. A knowledge of the relationship between the towed vehicle and the terrain profile will enable the design of an artificial road profile for fatigue testing with which similar stresses arise as during normal use. With the developed comparative method, a connection can be established between stochastic road profiles and road profiles containing artificially built obstacles.     

A free gait algorithm for quadrupedal walking machines

A free gait is a computer generated, rule-based gait for a walking machine to walk on rough terrain. Based on a given terrain map, the gait algorithm selects footholds for leg placements and determines the movements of legs and body. In the past, a few free gaits for hexapods have been developed. For quadrupeds, the only report on free gait was briefly mentioned in a paper by Hirose [Int. J. Robotics Res., 3(2) (1984)]. In this paper, a free gait algorithm for a quadrupedal walking chair is developed. For quadrupeds, the stability margin is small due to a small number of legs and the choices of a leg to be lifted are limited. Hence, deadlock situations may occur quite often. Many special techniques are incorporated into the algorithm in order to reduce deadlocks. This free gait algorithm adopts the wave-crab gaits as the primary gait because they are periodic and can provide good stability. The algorithm also adopts a non-periodic free gait to handle terrain with higher concentration of forbidden areas. This algorithm is evaluated under different terrain conditions using computer simulations. The results show that the performance is satisfactory on randomly generated rough terrain and needs improvement on manually generated rough terrain.