The panel met on December 10-11, 2018, at the National Academies of Sciences, Engineering, and Medicine (National Academies) facility in Washington, D.C., to review the In-House Laboratory Independent Research (ILIR) program projects in computational sciences conducted in 2018 at the following U.S. Army Research, Development, and Engineering Centers (RDECs): Armament Research, Development, and Engineering Center (ARDEC); Communications–Electronics Research, Development, and Engineering Center (CERDEC); and Tank Automotive Research, Development, and Engineering Center (TARDEC). The panel received overview presentations on the ILIR programs at each RDEC and technical presentations describing the projects. During each presentation the panel engaged in question-and-answer sessions with the presenter, and a general discussion with RDEC staff after the panel had formulated initial impressions and developed additional questions during its closed-session deliberations, conducted after the RDEC staff had concluded their presentations.
Project: Heterogeneous Visual Perception for Collaborative Localization of Autonomous Vehicle Teams
This project addresses the problem of coordinated perception for a team of robotic agents. The team members may have different capabilities and sensor modalities. They may also be at different locations and may be assigned different tasks. The objective of the agents is to coordinate, localize with respect of the environment and the rest of the team, and decide who is going to do what and at what time. These are decision and perception problems that are greatly affected by the uncertainty in the measurements. The key objective of the project is to quantify how pose uncertainty will affect the decision making of the agents. As a particular example, a small team with just two agents (one unmanned ground vehicle
[UGV] and one unmanned aerial vehicle [UAV]) is considered to demonstrate how pose uncertainty affects the decisions made by the team. It is assumed that the UAV has an on-board camera that can be used to detect the UGV. Similarly, the UGV has a stereo camera that can be used to detect the environment and the UAV flying overhead. The purpose of this illustrating example is to answer the following question: What is the best path for the UGV to reach the goal location?
The equations of motion have been derived for this setup, and several experiments have been conducted using a VICON and an Optitrack motion capture system, as well as using Gazebo, a virtual simulation and animation tool that comes with a Robot Operating System (ROS). The experimental and numerical simulation results correlate well with each other. The results to date show that the UAV/UGV team successfully coordinates to avoid obstacles. A paper summarizing this work has been submitted to the Institute of Electrical and Electronics Engineers’ International Conference on Robotics and Automation (IEEE ICRA) 2019 conference, and the acceptance decision is pending.
This project falls under the class of the so-called multi-robot simultaneous localization and mapping (multi-SLAM) problems, which is an active and important area of research in the multi-agent robotic systems community. As the current trend in the community is the design and operation of decentralized systems, the problem of coordination of the various agents in the team becomes increasingly relevant. A large body of literature already exists on this topic. One of the major challenges in such multi-agent scenarios is the so-called registration problem, namely, the process by which the various agents agree on which objects and features are observed by the different agents. From the presentation during the review, it was not clear whether this problem was considered or whether the registration is assumed to be given a priori. If feature registration is assumed to be solved, the rest of the problem is similar to a sensor-fusion problem, which represents a much easier problem to address.
This project aims at developing a multi-SLAM capability within ARDEC. Although the approach followed is standard, the principal investigator (PI) seems to be aware of the state of the art. This project was not renewed for FY19 owing to slow progress toward the goal; however, ARDEC provided alternative funding to support this research. Given the immense interest in multi-agent perception and the current level of research conducted at several academic institutions, industry, and other government laboratories, the competition is stiff. It is therefore understandable that it is not easy to produce new, publishable results in a short period of time in this area. The PI has proposed to extend this work in collaboration with the United States Military Academy (USMA) in the form of a team-on-team competition (e.g., Army versus Navy). If materialized, such a competition can serve as a great venue to engage USMA students in studying multi-agent tactics in the battlefield. In summary, this is a very timely and relevant topic, but more researchers and in-depth research are needed to generate results that rival the current state of the art.
Project: Numerical Shape Optimization Methods for Controlled Stress Distributions at Surface Interfaces
This research was devised to develop a finite element (FE)-based methodology for optimizing the shapes of interfaces between solid bodies in contact. The basic hypothesis for the work is that uniformity of the stress states in contacting bodies is a desirable design objective: stress uniformity will lead to lower peak values of stress than might otherwise occur and could possibly lead to localized material failure. The numerical method used a commercially available FE package, Abaqus, for the stress analysis, together with specialized algorithms or scripts written by the investigator using MATLAB and Python to alter the interface shapes. An iterative process was used to adjust the positions of nodal points along the boundaries of contacting bodies using a zero-order optimization method. Additionally,
a specialized nodal smoothing algorithm was developed to adjust the positions of interior nodal points for the Abaqus model. A multiobjective optimization process was also developed to allow for incorporation of weighted averages of selected stress components and strain energy densities in the numerical model. Analytical solutions are available for simple single-component problems (i.e., bodies without interfaces), and the numerical method successfully reproduced these. The method produced smoother interfaces than those predicted by commercial software and also showed favorable comparisons with numerical results reported by other investigators.
The work performed to date is promising and represents an innovative approach to adapting a commercially available software package for a special purpose problem of significant interest to the Army. The investigator has made good efforts to publish the work in the open literature and to present the findings at technical meetings; these are very important components of the peer-assessment process. It is too soon to determine the full impact of the work, but there are possible avenues for taking these basic research results into several application areas.
As an extension of the work, it would be worthwhile to consider the inclusion of a more elaborate gradient-based optimization method to provide more accurate values of the design parameters. Another possible extension of the research would be to include a capability for modeling frictional debonding, slip, and separation of interfaces as might occur, for example, along a reinforcing bar in a concrete structure.
Project: Deep Learning with Known Confidence
Confidence in the predictive capabilities of machine learning systems is mostly derived from tests performed with testing exemplars or based on historical performance over a period of time. A probabilistic perspective would provide a more meaningful prediction confidence and is the motivation behind this research project. The approach combines variational inference methods with statistical metrics to develop a stochastic variational inference approach for supervised deep learning neural networks. Two specific aspects of prediction capability—namely, prediction consistency (i.e., variability in predictions) and prediction pedigree (i.e., causality of input parameters)—are considered important targets for determination of confidence. The approach is developed in the context of classification problems.
Laboratory experimentation was performed with available data sets for the training and assessment of classification algorithms. Well-known machine learning libraries were deployed in the research, and the Army’s high-performance computing (HPC) environment was used for network training purposes. Baseline results for the work were obtained using traditional classifier methods. A stochastic dropout approach was deployed to develop the predictive distribution, and well-known statistical tests were then applied to assess the predictive distributions. The researcher demonstrated awareness of literature in the area and followed sound research practices, investigating the sensitivity of results on parameters such as network architecture and exploring the robustness of predictive capabilities through characterization tests across multiple data sets.
Meaningful results have been realized including a validation of the hypothesis that dropout-based variational inference can be used to build a meaningful predictive distribution. Results show that the mean obtained from a prediction distribution mean is more informative about confidence than a single prediction; the variance, however, does not really identify low-confidence predictions. The confidence estimation metrics were found to be sensitive to the choice of loss function used during the network training process
and were not always successful in identifying missed classifications. Work on prediction pedigree is still in progress, but early indications are that the pedigree technique can create a meaningful saliency map of the image to be classified. The PI proposes to expand the use of this approach in classifying radio frequency (RF) signal patterns. The time-dependent nature of this data will introduce an additional degree of research challenge to this problem.
This is a solid basic research project that is in the sphere of work being pursued under the Defense Advanced Research Projects Agency’s Explainable Artificial Intelligence (DARPA XAI) program. Confidence in decisions will become increasingly important as the Army seeks to embrace automated approaches in a variety of functions. The lack of trust in automated processes has typically relegated decisions from machine learning algorithms to noncritical functions. As machine learning algorithms become more pervasive in military applications, this research will help provide a more rational basis for use in more critical functions. It is also noteworthy that the Army has access to a combination of unique data sets and computing resources that will contribute to the success of this work. The PI is supported at a level of 30 percent of his time on this project; more significant engagement would certainly contribute to accelerated progress. The PI seems well connected with the community pursuing similar work and has sought feedback from members of that technical community. He has presented in multiple technical exchanges and is currently preparing a publication for a conference.
Project: Behavior-Based Convoy Formation Control for Multi-Robot Teams
A centralized approach to convoy formation is known to be both computationally demanding and susceptible to a single point failure. The research project seeks a robust decentralized approach to this problem that requires no shared communication among the vehicles. The goal is to develop a methodology that would not only be robust to system errors and communication loss but would also be highly scalable. The research focuses on the adaptation of behavior-based formation control deployed in robotics as a solution to the convoy formation problem. A convoy manager generates waypoints on the basis of mission requirements, and an execution layer combines predefined behaviors in conjunction with waypoint information to develop a series of convoy movements. The approach is implemented on a simulation platform called an Autonomous Navigation Virtual Environment Laboratory (ANVEL), which has the capability to include data from the Global Positioning System (GPS), radar, and Light Detection and Ranging (LIDAR) sensors. The approach has been tested for typical cases of moving obstacles getting between the leader and followers and shows how following vehicles are able to maneuver around obstacles and return to following the leader. Additional fine-tuning of an approach to introduce a constraint on the separation between leader and follower may be necessary for certain tight turning paths.
Under the TARDEC Ground Vehicle Robotics (GVR) Programs of Record, logistic resupply represents an important focal area. In this thrust, there is an explicit emphasis on autonomous convoy operations, including the use of the leader/follower approach. The research project, therefore, is highly relevant and aligned to the mission of the organization. The researchers are aware of prior and ongoing work in this area and are examining these in the context of the proposed solutions. They are seeking to evaluate the proposed approach for typical scenarios that an Army convoy may encounter. The ANVEL platform provides a convenient testbed to assess the proposed methodology. This is not, however, without its own challenges. Vehicle Stability System Plus (VSS+), a widely deployed tool at TARDEC that interfaces with ANVEL for the simulation and testing of convoy-based unmanned ground vehicle systems, has convoy communications as a core feature. This core feature needs to be modified to accommodate the
decentralized approach that is being pursued in the current work. The lack of backward compatibility in newer versions of the ANVEL platform is another undesirable element that has resulted in slower progress. Continuing work on the project seeks to explore reinforcement learning as an approach to develop other behavior models for use in the proposed methodology. It was not made clear from the presentation during the review how this would be implemented or what additional information was being sought in the alternative behavior models.
The research falls within the broad category of work being directed toward self-driving and autonomous vehicles. While the convoy formation aspect is of special interest to ground mobility logistics, the work is also relevant to the deployment and movement of autonomous robotic vehicles in the battlefield. The approach and the results obtained to date are promising. Among the scenarios used to test the robustness of the proposed methodology, due consideration needs to be placed on examining the impact of losing or failing to identify one or more waypoints on mission success. Other metrics that will be used to gauge the effectiveness of the proposed approach need to be explicitly defined. The level of funding supports 25 percent of the co-PIs’ time. This level of staffing is able to support an exploratory effort, but more staffing would be needed to transition the work into field deployment. The PIs have clearly outlined the expected deliverables for the research effort; these are in the form of operating system agnostic software for the ANVEL simulation platform. At the time of the briefing, a conference publication related to the research was in preparation.
Project: Deep Learning for Autonomous Driving
This project investigates the problem of establishing trust within a team of homogeneous autonomous agents. The key objective is to leverage different agents’ experiences online in order to obtain an aggregated model of trust across all agents in the team. Starting from a previous computational trust model (RoboTrust), the PI proposes to account for nonlinearities using deep neural networks. The new trust model, called NeuroTrust, can adapt to a stochastic environment through reinforcement learning.
The specific problem addressed is convoy controller design. Specifically, the objective is to design a cruise control to maintain the desired distance between the vehicles. It is assumed that the speed of each vehicle depends on the level of trust, and hence the follow distance varies with the level of trust. The objective is for the vehicles to travel as fast and as close as possible, while hitting as few obstacles as possible. In the absence of actual data, simulated convoy scenarios are used to compare RoboTrust and NeuroTrust. The PI makes use of available open source software such as MATLAB, ANVEL, and deep learning (DL) software toolkits.
Quantifying trust between autonomous agents and humans is a challenging problem. Many groups, especially in the human factors and human–machine interaction communities, have investigated this problem with varying levels of success. The proposed approach uses a binary history of observations to compute trust. It builds on an existing static model of trust (RoboTrust), but it extends it to account for nonlinearity, using a deep neural network architecture, which is quite common to account for unknown dynamics and nonlinearities in a nonparametric fashion. Since no explicit model is assumed, reinforcement learning ideas are used to update the control policy based on the current level of trust.
This research supports autonomous driving behaviors for ground vehicles, which is one of the critical aspects of the U.S. Army Training and Doctrine Command’s (TRADOC’s) Intelligent Mobility initiative and TARDEC’s strategic efforts in autonomy-enabled systems. Although the effort focuses on the convoy problem, the issue of trust is crucial for any manned-unmanned teaming (MUMT) scenario.
This project constitutes basic research on an important topic that is of great interest in the robotics community in general. The PI is well aware of the current state of the art. One of the issues that was not
addressed in this project (and would be a good area for future research) is a more credible validation of RoboTrust against human studies. In addition, a sensitivity study of the RoboTrust model in terms of the confirmation and tolerance parameters is needed to ensure that the results from the simulations are consistent and repeatable for a wide range of realistic situations. An additional challenge for real-world implementation is the assumption of an infinite number of iterations and the availability of a very large data set to converge to an optimal model of trust.
Project: Information Reliability—Identification of Drivers and Quantification of Their Contributions
This research project aims at answering the following questions: (1) What information is important and/or relevant for the reliable operation of unmanned, autonomous ground vehicles? (2) How can one validate and test autonomous vehicles prior to deployment in the field? and (3) How can one design a decision-making pipeline that is transparent and interpretable to the human operator? The basic idea applied to answer these questions is to collect data from several field tests (as part of the Expedited Warfighter Experiment) and analyze the data using a Bayesian network.
Bayesian networks are graphical models that are able to encode causal relationships between events. The structure of the network is very important to get credible results. It is proposed to use domain expertise to find the structure of the network prior to training, which is the common practice. Preliminary results show that this approach is able to predict the major effects that cause a driver to take control of an autonomous vehicle in a military convoy situation.
This research project addresses some very important and urgent questions in the area of autonomous systems. Currently, the workhorses behind most autonomous systems are perception and decision-making algorithms based on deep neural networks. These networks have shown impressive results in many challenging pattern recognition problems that may elude human observers. Despite their great success, however, these algorithms are somewhat brittle, and they suffer from unpredictability and childlike behavior. They confuse correlation for causality and often find patterns where none exist, becoming confused and perplexed, unable to come up with the correct answer even in very simple scenarios where the answer would have been clear to a human observer. This discrepancy between theoretical statistical accuracy and demonstrated fragility in practice makes these algorithms currently unsuitable for many safety-critical systems. Furthermore, their reliance on high-quality data makes these algorithms questionable for military applications, where data and prior experience are incomplete or nonexistent. Given this context, this research project asks the right questions, albeit in a rather limited context. A complete answer to these questions is the Holy Grail in the area of autonomy, and the fact that the PI uses a combination of ideas from control, communication, and information theory probably indicates the right path.
The topic addressed in this project is timely; therefore, it comes as no surprise that it is a very active area of research by many researchers in industry and academia alike. Nonetheless, it differs from current work in the area of commercial self-driving vehicles in several aspects. The latter assumes a structured environment (e.g., highway driving) with additional help from infrastructure (e.g., GPS and vehicle-to-vehicle [V2X]) and the availability of a large amount of data. This is not the case with military autonomous ground vehicle applications. New techniques and methodologies have to be developed. The project is a good step toward this goal. Given the sound problem formulation, there is a good opportunity here to go beyond the example of human driver intervention in an automated convoy scenario and try to extend the approach to answer more fundamental questions relating to the key question originally posed—namely, what information is important and/or relevant to process for the task at hand? In terms of the core methodological approach,
the possibility of generating the structure of the Bayesian network from data (and in doing so bypassing the need for human expert intervention) is worth investigating.
Project: The Development of an Optimized Computational Framework for Multiscale Simulations
The objective of this project is to develop an optimized multiscale framework that couples finite elements methods (FEM) and discrete elements methods (DEM) to solve real-world large vehicle mobility models such as vehicle dynamics soil models. The technical approach is to couple FEM and DEM fully integrated within a monolithic flexible multibody dynamics solver. The mathematical framework is based on the hierarchical multiscale simulation (HMS) method developed at the Army Research Laboratory (ARL). The HMS includes a stand-alone module to facilitate multiscale model evaluation that is responsible for the creation, scheduling, and control of nested applications. Features of the HMS also include the incorporation of lower-scale models without modifications to existing codes and allow for multiple evaluations of the lower-scale model at once. Consequently, the approach optimizes the models and framework and applies the developed multiscale method for full vehicle mobility simulations. It is expected that the co-simulation of full vehicle models with the developed multiscale finite element-discrete element (FE-DE) method will be completed in 2019. Future directions include research on uncertainty quantification analysis to investigate the robustness of multiscale simulation.
The results obtained within this project show that the predictions of the coupled multiscale FE-DE model simulation agree with available experiments. They also show that the multiscale FE-DE model significantly improves computational time over the pure DE model and that the HMS multiscale model exhibits good scalability.
This is a project of high technical merit and of fundamental interest that appropriately integrates theory; modeling and simulation; and validating experimental efforts. The basic hypotheses of the work have been validated, namely that (1) the developed multiscale off-road mobility model runs significantly faster than the similar pure DE model, (2) the ARL-developed multiscale framework ensures more efficient parallel computing scalability for large-scale simulation models in the context of off-road mobility simulation, and (3) the developed multiscale tire–soil interaction model allows for running the tire–soil interaction simulations with reasonable computational time.
The integration of FE and DE methods within the ARL-developed HMS framework appears to be innovative and builds on ARL’s strong capabilities in multiscale materials simulations. Specifically, the multiscale off-road mobility simulation capability is being integrated with the ARL’s HMS scale-bridging algorithm to enhance the computational load-balancing ability for multiple lower-scale models.
The project has resulted in two publications (one conference proceeding and one manuscript under review) and two conference presentations. The project includes external collaborators and is a collaborative project among the University of Iowa, TARDEC, and ARL. The investigators are cognizant of similar work performed at other institutions and have built on this body of knowledge by coupling vehicle multibody dynamics and optimized tire–soil interaction modeling. The ILIR project aims to bring together ongoing collaborative efforts to develop an enhanced multiscale off-road mobility solver by leveraging the Department of Defense’s (DoD’s) Supercomputing Resource Center (DSRC) high-performance computing (HPC) resources. The technical work is competitive with other efforts in this area within the scientific community.
This project supports TARDEC’s mission to develop, integrate, and sustain the right technology solutions for all manned and unmanned Army ground systems and off-road mobility. Because the hierarchical FE-DE multiscale model developed under the auspices of this ILIR project can be broadly
applicable to other areas beyond tire–soil interaction simulations, there exist several opportunities to apply these simulation capabilities to other DoD mission areas in support of broad DoD mission needs.
Beyond the technical challenges associated with the development of a predictive hierarchical FE-DE multiscale tire–soil interaction simulation capability, challenges associated with project execution also exist. These include, but are not limited to, access to appropriate computing resources and limited staffing to complete the tasks associated with the project.
TARDEC Crosscutting Findings
The development of advanced simulation capabilities to predict the behavior of systems and components in extreme environments cuts across virtually all Army mission areas and is pervasive across all mission needs. Consequently, efforts in all areas of computational science, including multiscale materials simulations, deep learning algorithms, multi-agent optimization, and vehicle autonomy, all of which are widely applicable to the Army mission, need to be supported across basic and applied research programs, across organizations such as ARL and ARDEC. The ILIR program is an important source of support for such programs that link basic and applied research efforts. ILIR resources can help enable the integration and coordination of the Army’s various simulation efforts in the areas of systems in extreme environments, toward the development of an integrated and comprehensive predictive simulation capability.