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.     

工程场地的特征与工程地质分区

工程场地与城市在空间尺度上相去甚远 ,但工程场地所具备的与环境行为相近的共性 ,使其属性分析可以推广应用到城市建设所面对的整个工程地质环境。从这一意义上说 ,工程场地是整个城市宏观工程地质环境的微观单元 ,众多微观单元的组合构成了环境的宏观总体。为此 ,工程场地的三大属性———特性、适宜性、稳定性 ,可以作为城市工程地质环境分析中分析方法的基础。从工程场地分析推广到城市工程地质环境分析 ,还将有助于保证城市工程地质环境分析的定量化水平。     

Maneuver analysis methodology to predict vehicle impacts on training lands

Tactical mobility analysis techniques were merged with land management strategies to assess potential impacts of vehicle operations on training areas for rangeland planning and management. A vehicle mobility analysis was performed for a suite of vehicle types using the NATO Reference Mobility Model (NRMM II). Input parameters include terrain information (soil type, slope, vegetation, surface roughness, soil strength), terrain surface condition based on climate (terrain strength, freeze–thaw, moisture content, snow cover), and vehicle specifications (tire, power train, weight on each axle, ground clearance, dimensions, ride). The vehicle performance was spatially mapped over the terrain for different seasons of the year and used to calculate the maneuverable acreage, which was compared to acreage needed for training requirements. This can be related to land capability based on expected training impact (Maneuver Impact Miles, MIM) and Land Condition Curves which link training density to land condition. This methodology can be used to determine the suitability of training lands and the degree of land management or rehabilitation expected. The methodology was applied to the transformation of the Alaska training lands to support a new brigade unit called the Stryker Brigade Combat Team (SBCT3), but is equally useful for other training areas and military units. For summer use, Alaska training lands are capable of supporting four times the projected training requirements. For winter, when the ground is frozen, more than 10 times the area needed was available.     

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.     

Quantifying vegetation biomass impacts on vehicle mobility

Soil impacts on vehicle mobility are well known; however, most data are for bare soil or the type and amount of vegetation is not documented. This study summarizes results from experiments to quantify the effect of above ground and below ground vegetation biomass on vehicle performance. Soil–vegetation combinations of three soils and three grasses were used. The vegetation was tested at various growth stages and was also subjected to stressors such as trafficking, burning, and cutting. Vegetation measurements included above ground (leaves and shoots) and below ground (root) biomass weights, lengths, diameters and surface area parameters. The soils were characterized for size distribution, moisture, density and terrain strength for each test condition. Vehicle traction and motion resistance were measured for each soil–grass combination using the CRREL Instrumented Vehicle. Results showed an increase in net traction biomass in sandy soils. For clay soils above ground biomass generally increased resistance while increased root diameter clearly decreased resistance. This study represents the first measurements quantifying the impacts of specific biomass parameters on vehicle mobility. The results will serve to guide new experimental methods, improve datasets, and develop physics-based models for years to come.     

Automated antenna positioning algorithms for wireless fixed-access networks

This article addresses a real-life problem - obtaining communication links between multiple base station sites, by positioning a minimal set of fixed-access relay antenna sites on a given terrain. Reducing the number of relay antenna sites is considered critical due to substantial installation and maintenance costs. Despite the significant cost saved by eliminating even a single antenna site, an inefficient manual approach is employed due to the computational complexity of the problem. From the theoretical point of view we show that this problem is not only NP hard, but also does not have a constant approximation. In this paper we suggest several alternative automated heuristics, relying on terrain preprocessing to find educated potential points for positioning relay stations. A large-scale computer-based experiment consisting of approximately 7,000 different scenarios was conducted. The quality of alternative solutions was compared by isolating and displaying factors that were found to affect the standard deviation of the solutions supplied by the tested heuristics. The results of the simulation based experiments show that the saving potential increases when more base stations are needed to be interconnected. The designs of a human expert were compared to the automatically generated solutions for a small subset of the experiment scenarios. Our studies indicate that for small networks (e.g., connecting up to ten base stations), the results obtained by human experts are adequate although they rarely exceed the quality of automated alternatives. However, the process of obtaining these results in comparison to automated heuristics is longer. In addition, when more base station sites need to be interconnected, the human approach is easily outperformed by our heuristics, both in terms of better results (fewer antennas) and in significant shorter calculation times.     

An integrated approach for earthwork allocation,sequencing and routing

Planning techniques for large scale earthworks have been considered in this article. To improve these activities a “block theoretic” approach was developed that provides an integrated solution consisting of an allocation of cuts to fills and a sequence of cuts and fills over time. It considers the constantly changing terrain by computing haulage routes dynamically. Consequently more realistic haulage costs are used in the decision making process. A digraph is utilised to describe the terrain surface which has been partitioned into uniform grids. It reflects the true state of the terrain, and is altered after each cut and fill. A shortest path algorithm is successively applied to calculate the cost of each haul, and these costs are summed over the entire sequence, to provide a total cost of haulage. To solve this integrated optimisation problem a variety of solution techniques were applied, including constructive algorithms, meta-heuristics and parallel programming. The extensive numerical investigations have successfully shown the applicability of our approach to real sized earthwork problems.     

Estimation of net traction for differential-steered wheeled robots

We present a method for estimating the net traction and resistive wheel torques for a suspensionless, differential-steered robot on rigid or deformable terrain. The method, based on extended Kalman-Bucy filtering (EKBF), determines time histories of net traction and resistive wheel torques and wheel slips during steady or transient maneuvers. This method assumes good knowledge of the vehicle dynamics and treats the unknown forces and moments due to terrain response as random variables to be estimated. A proprioceptive sensor suite renders a subset of the unknown forces and associated wheel slip and slip angles observable. This methodology decouples semi-empirical terramechanics models from the net effect of the vehicle-terrain interaction, namely the net traction developed by the vehicle on the terrain. By collecting sensor data and processing data off-line, force-slip characteristics are identified irrespective of the underlying terramechanics. These characteristics can in turn support development or validation of terramechanics models for the vehicle-terrain system. For autonomous robots, real-time estimates of force-slip characteristics can provide setpoints for traction and steering control, increasing vehicle performance, speed, and maneuverability. Finally, force-slip estimation is the first step in identifying terrain parameters during normal maneuvering. The methodology is demonstrated through both simulation and physical testing using a 13-kg robot.     

Predictability of boreal forest soil bearing capacity by machine learning

In forest harvesting, terrain trafficability is the key parameter needed for route planning. Advance knowledge of the soil bearing capacity is crucial for heavy machinery operations. Especially peatland areas can cause severe problems for harvesting operations and can result in increased costs. In addition to avoiding potential damage to the soil, route planning must also take into consideration the root damage to the remaining trees. In this paper we study the predictability of boreal soil load bearing capacity by using remote sensing data and field measurement data. We conduct our research by using both linear and nonlinear methods of machine learning. With the best prediction method, ridge regression, the results are promising with a C-index value higher than 0.68 up to 200 m prediction range from the closest point with known bearing capacity, the baseline value being 0.5. The load bearing classification of the soil resulted in 76% accuracy up to 60 m by using a multilayer perceptron method. The results indicate that there is a potential for production applications and that there is a great need for automatic real-time sensoring in order to produce applicable predictions.