5

Information Sciences: Autonomous Systems

INTRODUCTION

The Panel on Information Sciences at the Army Research Laboratory conducted its review of ARL’s Autonomous Systems program on August 13-15, 2013. This chapter provides an evaluation of that work, recognizing that it represents only a portion of ARL’s information sciences core technology competency portfolio.

While there was considerable variation in both the quality and impact of the research presented, the researchers were largely aware of the progress in their fields, and that had a noticeable impact on their own work. ARL has recently recruited a number of very promising early-career scientists. Careful attention needs to be directed at ensuring that they receive appropriate mentorship and career development opportunities as they develop their individual research portfolios. The new indoor Military Operations on Urban Terrain (MOUT) facility was impressive and will go a long way in furthering the goals of the intelligence and planning program. The tour of the ARL Sensors and Electron Devices Directorate’s Specialty Electronic Materials and Sensors Cleanroom Research facility helped the review team understand the infrastructure support available to ARL researchers.

A summary assessment of research in each of four subject areas—manipulation and mobility, perception, robotic intelligence, and human–robot interaction—is presented in the following sections of this chapter. ARL has a leading program in the area of small-scale robotics. A demonstrated ability to design, fabricate, and test these devices gives it a place of distinction in this field. Similarly, research in the area of perception is being performed at a high level. With a mission to develop machine understanding of objects, actions, and interrelationships in a specified environment, this work is critical for advancing the state of autonomous systems. Ongoing research is focused on advancing unsupervised approaches to human detection and advancing sensing and perception capabilities on constrained platforms. Research in the areas of human–robot interaction and intelligence is addressing important problems of mapping, cognition, and communication, as well as issues related to trust in autonomous systems. This research is cutting edge and comparable to work at federal, university and/or industrial laboratories here and abroad, and portions of the work are poised for successful transition to applied research.

ACCOMPLISHMENT AND ADVANCEMENTS

All elements of the autonomous systems research program at ARL have continued to show progress, both in the quality of work and dissemination of results in high-quality publications. The program focuses on mobility and manipulation of robotic devices and on technologies that improve the usefulness of these devices, such as intelligence, perception, and improved human–robot interaction. The ARL research program is part of a larger collaborative effort involving external partners. A better definition of the role of the internal research in the overall program goals and continued collaboration with partners is strongly encouraged.

Research in the area of manipulation and mobility is closely linked to the ARL’s Collaborative



The National Academies | 500 Fifth St. N.W. | Washington, D.C. 20001
Copyright © National Academy of Sciences. All rights reserved.
Terms of Use and Privacy Statement



Below are the first 10 and last 10 pages of uncorrected machine-read text (when available) of this chapter, followed by the top 30 algorithmically extracted key phrases from the chapter as a whole.
Intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text on the opening pages of each chapter. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.

Do not use for reproduction, copying, pasting, or reading; exclusively for search engines.

OCR for page 51
5 Information Sciences: Autonomous Systems INTRODUCTION The Panel on Information Sciences at the Army Research Laboratory conducted its review of ARL’s Autonomous Systems program on August 13-15, 2013. This chapter provides an evaluation of that work, recognizing that it represents only a portion of ARL’s information sciences core technology competency portfolio. While there was considerable variation in both the quality and impact of the research presented, the researchers were largely aware of the progress in their fields, and that had a noticeable impact on their own work. ARL has recently recruited a number of very promising early-career scientists. Careful attention needs to be directed at ensuring that they receive appropriate mentorship and career development opportunities as they develop their individual research portfolios. The new indoor Military Operations on Urban Terrain (MOUT) facility was impressive and will go a long way in furthering the goals of the intelligence and planning program. The tour of the ARL Sensors and Electron Devices Directorate’s Specialty Electronic Materials and Sensors Cleanroom Research facility helped the review team understand the infrastructure support available to ARL researchers. A summary assessment of research in each of four subject areas—manipulation and mobility, perception, robotic intelligence, and human–robot interaction—is presented in the following sections of this chapter. ARL has a leading program in the area of small-scale robotics. A demonstrated ability to design, fabricate, and test these devices gives it a place of distinction in this field. Similarly, research in the area of perception is being performed at a high level. With a mission to develop machine understanding of objects, actions, and interrelationships in a specified environment, this work is critical for advancing the state of autonomous systems. Ongoing research is focused on advancing unsupervised approaches to human detection and advancing sensing and perception capabilities on constrained platforms. Research in the areas of human–robot interaction and intelligence is addressing important problems of mapping, cognition, and communication, as well as issues related to trust in autonomous systems. This research is cutting edge and comparable to work at federal, university and/or industrial laboratories here and abroad, and portions of the work are poised for successful transition to applied research. ACCOMPLISHMENT AND ADVANCEMENTS All elements of the autonomous systems research program at ARL have continued to show progress, both in the quality of work and dissemination of results in high-quality publications. The program focuses on mobility and manipulation of robotic devices and on technologies that improve the usefulness of these devices, such as intelligence, perception, and improved human–robot interaction. The ARL research program is part of a larger collaborative effort involving external partners. A better definition of the role of the internal research in the overall program goals and continued collaboration with partners is strongly encouraged. Research in the area of manipulation and mobility is closely linked to the ARL’s Collaborative 51

OCR for page 51
Technology Alliances (CTAs) in Autonomous Systems, 1 where significant collaboration with those partners is to be found. Three areas were highlighted during the review: replicating locomotion found in biological systems to improve robot mobility, autonomous manipulation of robots, and piezomicroelectromechanical systems (piezoMEMS) technologies to develop small-scale robotic systems. New updates on the project related to the CANID robot, 2 mimicking the movements of a canine hound, were presented. The primary thrust of the new development was the addition of a flexible spine to the robot. This is a challenging and interesting project, and the addition of this key degree of freedom to a walking robot is a good idea, because it more closely resembles the complexity found in nature. There needs to be a concerted effort to better understand the physics of this machine. Researchers have proposed low-order models for the system, and it might be fruitful to continue this line of inquiry. Good modeling will be imperative in any numerical simulations required to explore gains that are possible and to guide the focus of new experimentation. Work related to self-righting robots is of a very high caliber and also has direct applications in the field. This was evident from the fact that the project was conceived through interactions with soldiers who are often confronted with the task of retrieving immobilized robots in combat. The research seeks to develop solutions for a broad class of physical conditions that affect stability of mobile robotic platforms. The current focus is on examining the underlying mechanical issues of self-righting in a quasi-static environment. It was not clear how the upright, stable position would be sensed on a sloping surface, where self-righting is most likely needed. ARL’s intention to move toward a consideration of dynamics in 2014 is applauded. The piezoMEMS research and associated small robotics effort, in the Board's judgment, is first rate, with elements that are at the vanguard of this field. The robotic devices under development with integrated piezoelectric materials demonstrated work that is at the forefront of MEMS design, fabrication, and experimentation. Specifically, the work in motion generation at the MEMS scale is seminal. Large- amplitude motions are being created at the micron scale using integrated actuators, structures, and electronics, co-fabricated on silicon. Techniques and approaches to generating articulating limbs with integrated flexure hinges and actuators represent advances in the engineering of MEMS technology. This has broad implications and applications to numerous MEMS systems—for example, MEMS-based microscale sensors and instrumentations such as mass spectrometers on a chip, drug delivery systems, and chemical assay analysis, where controlling microfluids are of fundamental importance. The work being performed by ARL in the piezoelectric actuation of MEMS will impact more than just the creation of bioinspired microscale robotic systems. The research projects in perception are of a high caliber. The work focuses on developing techniques that allow for developing a description of the robot’s environments from sensor data. While there has been considerable progress toward describing environment for the purpose of mobility, deriving higher-level descriptions such as subtle cues and references that distinguish different behaviors and intents, recognition of specific classes of objects and features that are directly relevant to tactical behaviors, and labeling of object, features, and terrain classes remain a challenge. The current research plan is focused on three critical areas: perception on constrained platforms, robotic intelligence, and human–robot interaction. Within the area of sensing and perception on constrained platforms, the scale and size of platforms being explored in the autonomous systems enterprise pose technical challenges in sensor design. Sensors have to deliver the requisite accuracy and precision for surveillance and navigation, but they also have to reconcile with power limitations on smaller platforms. Other problems being addressed in the perception area include human detection in still images and strengthening object and material recognition capabilities, including an ability to recognize actions and imminent actions. The latter is based on scene parsing and action grammar and represents an 1 There are currently two active CTAs related to autonomous systems: Micro Autonomous Systems and Technology (MAST) and Robotics. 2 CANID is a quadruped designed to test hypotheses regarding dynamic bounding using an actuated compliant spine mechanism. 52

OCR for page 51
interesting approach. The problem of developing real-time human detection algorithms is important, because existing approaches are based on supervised learning techniques that are computationally cumbersome. In operational environments that may be diverse and exhibit large variations, computational efficiency is an important consideration. The unsupervised learning approach used in this work is clearly more efficient but may still yield false positives, and additional work is required to overcome this drawback. The basic approach for each of the research tasks is fundamentally sound, breaking new ground. Results of this research are publishable in archival literature, and the work is at par with research being done at universities and other laboratories. The researchers are aware of related work being done elsewhere and recognize deficiencies in their individual approaches. Each presented a good plan for the future activities. The primary accomplishments in robotic intelligence are advances in mapping capability, control for communications, and cognition. Much of this work is being published in top journal and conference venues (including the International Conference on Robotics and Automation, the Institute of Electrical and Electronics Engineers Proceedings, and the International Journal of Robotics Research), which attests to the overall quality of the research. Former students funded by the CTAs have been recruited to ARL and are important contributors to the research effort. Some of the 6.1 (basic) research projects are also making their way to 6.2 (applied research) applications, an important step for transitioning this work to the field. In the area of long-duration three-dimensional (3D) mapping and navigation, the principal focus is on the development of a laser-based approach to 3D mapping that combines features from three existing mapping techniques. Demonstrating the effectiveness of the approach by deploying it on physical robots is an important accomplishment. Another thrust of intelligence research is on developing robust methods for control of mobility and communications. The primary focus of this effort is the development of a centralized, optimal solution of mobile node positions, subject to point-to-point communications constraints. While promising results were presented, additional details are necessary, particularly for issues such as scalability and whether such issues impose limitations on the proposed approach. Research in intelligence is also examining new cognitive architectures for robotic control. This work is focused on developing a new cognitive architecture that combines long-term memory, working memory, and perception. The mapping problem is driving this development, but it is difficult to understand how the cognitive approach improves mapping performance. In the area of human–robot interaction (HRI), research at ARL is looking at design issues for safe operation of autonomous reconnaissance systems in complex environments. The emphasis of this effort is human factors experiments to investigate interaction with, and control of, multiple autonomous systems. The design of interfaces is an important aspect of this investigation. Studies are focused on graphical user interfaces (GUIs), multimodal interfaces (including voice), and telepresence with stereovision and haptic interfaces. The experimentation conducted at Fort Benning has yielded an important basis for making design decisions. For example, experiments have demonstrated voice commands to be suitable for discrete actions but less so for controlling continuous processes. Similarly, the research has demonstrated how audio cues in 3D improve situation awareness in telepresence tasks. The RoboLeader project continues to be an important component of the ARL HRI program. The research draws on a large body of experimental work related to evaluating the effectiveness of autonomous and semiautonomous control of teams of robots. The RoboLeader intelligent system was evaluated in this context and shown to provide benefits from the standpoint of both task performance and workload management. Testing with humans showed individual differences in performance: People with high spatial ability and significant videogame experience had better situation awareness of the mission environment. The results have implications for personnel selection, training, and user interface design. Another research thrust in the HRI arena is bringing greater understanding of automation actions to the human in the loop. The focus of this effort is the use of visual display screen overlays to communicate robot perceptions and intentions to a human operator during an automated navigation task. The experimental approach is sound and based on prior studies of shared mental models and automation transparency. Results of this work support the use of such visual aids as an approach to reduce 53

OCR for page 51
teleoperation occurrences, teleoperation times, and subjective workload. The HRI group’s publications reflect a broad understanding of the science and research conducted elsewhere—related work is cited and contributions are placed in context. The group has also edited a book, Human-Robot Interactions in Future Military Operation, 3 which includes input from external sources, including academia. It also has a demonstrated record of successfully transitioning its work to the U.S. Army Tank Automotive Research Development and Engineering Center (TARDEC) robotics programs. OPPORTUNITIES AND CHALLENGES It was not clear how the individual research projects in each area—manipulation and mobility, perception, robotic intelligence, and human–robot interaction—of the ARL autonomous systems enterprise fit within the larger research effort. Without such a road map, there is very little indication of the connectivity of the research projects, both within the subareas and across the enterprise. At a fundamental level, ARL can take additional steps to enhance the quality and impact of its research efforts. There is a trinity in research and development: analysis, computation or simulation, and testing. Analysis is essential, and there is room for improvement on this front, before one proceeds with numerical simulation or building and testing artifacts. Results from analytical modeling can guide the subsequent steps in development and identify possible missteps: This analytical component needs to be integrated into the approach to research. For example, it is not enough to build a robot and begin to generate a gait similar to that of an animal—one must understand the physics behind the energy converted in a machine when using such a method of locomotion. In the area of manipulation and mobility, there is need for a more coherent approach to vehicle design and development. It would appear that while many excellent issues are being addressed, the overall approach is somewhat ad hoc. For instance, there is an absence of nondimensional scaling in platform design. Characterization of the fundamental physics of flying vehicles—length-to-diameter (L/D) ratio, drag polar, coefficients of lift and drag, and power requirements—must be part of the basic design philosophy. Similarly, metrics for system performance evaluation in generalized terms, such as actuation efficiency, propulsive efficiency, hover power loading, power to weight ratio for the actuator, endurance, and specific energy of the fuel source, would add focus to coupled systems and vehicles to include physics-based performance attributes. For robot systems, the specific resistance or cost of transport for any locomotive machine, natural or man-made, is a measure of a machine’s locomotive efficiency. It would be beneficial for ARL to encourage this traditional thinking as part of the research mindset. There is an opportunity to perform simulations of robots and vehicles based on analytical models of the physical systems operating in different environments and to include uncertainties in these models. The models can further be coupled to real control systems, leading to hardware-in-the-loop control design. The ARL may need to consider procuring development hardware, such as D-Space, for robotic controller development. Such systems would accelerate results and allow the integration of complex, nonlinear controllers, based on traditional sensors and sensor fusion, states of the machine, learned behaviors, and complex logic of state machines. Once developed, successful controllers could be programmed, at which point the developmental kinks have be sorted out, to perform laboratory and field tests of the new controllers. Integration of systems components is essential to the robotics research area. The system is much more than the sum of its components—nonlinear interactions, sometimes stochastic in nature, can have significant impact on overall operation. In this context, an integrated approach (systems engineering) is a fundamental (6.1 level) domain of research. Such an approach will also allow researchers to best trade 3 Michael Barnes and Florian Jentsch, eds., 2010, Human-Robot Interactions in Future Military Operations, Ashgate Publishing Company, Burlington, Vt. 54

OCR for page 51
concept options against the desired output or value functions. There is an opportunity for the ARL autonomous systems enterprise to assume the leadership in advancing a highly quantitative and scientific approach to systems engineering as it relates to integration of systems components into the robotics research area. Other general suggestions related to this area of effort may be summarized as follows: • Notionally establish a family of small robots (ground and air) of varying size (between 1.0 g and 100 kg) and define and reach both the baseline and performance goals for each robotic class. These specifications could be based on a limited selection of potential Army scenarios or vignettes. • Establish a directed robotic mobility propulsion effort to unify and direct activities required to produce very high power and high energy density systems. • Establish an integrated design and optimization methodology (considering key parameters like energy, power conversion efficiency, and locomotive efficiency) for the design of these highly integrated robotic systems. The interdependence of all the subsystems will quickly become clear to the researchers as they try to categorize future robotic systems. • Consider the establishment of a robotic mobility systems integration laboratory. This laboratory would allow for the integration of complete physics-based air and ground mechanics models of selective robots with candidate control systems in a simulated real-world operating environment. Although the mission statement pertaining to perception research is rather broad, the actual ongoing focus is restricted to a rather narrow set of problems. It was not clear if this focus was driven by gaps or deficiencies identified by the Army or how this work fits with contributions from partners. As an example, perception needs to support more than obstacle avoidance; it needs to support a richer semantic understanding of terrain that would be useful for people as well as autonomous vehicles. Furthermore, there is a need to address scalability issues with regard to sensor capabilities, such as varying fidelity and power requirements with size. Sensors provide measurements of data pertinent to the operational environment. For autonomous behavior, however, it is necessary to process these physical measurements to glean information. In the area of unsupervised human detection, researchers need to explore a hybrid supervised/unsupervised algorithm that would not only be computationally efficient but could also curtail the number of false positives that the current approach seems to yield. Sensors are also linked to communications. The processing and communications power available to a single platform determines what is transmitted—measurements, processed data, or commands. The work on parsing and action grammars could benefit from transmitting unknown constructed images over a communication network to a node with greater processing capability and reference data. In this context, focus needs to be directed at combining scene parsing, scene surface layout analysis, and 3D reconstruction to advance the state of the art in overall scene understanding. It may be useful to place bounds on the problem dictated by mission requirements. This would help identify quantitative metrics against which progress in research tasks can be measured. Ongoing research in perception is aimed at enabling cooperative interaction between robots and humans at multiple levels. Accomplishing this within a mission context, accepted military doctrine, and social norms of the society in which the soldier-robot teams operate is a major technical challenge. Overall, the research programs in intelligence reflect a broad understanding of the underlying science and research conducted elsewhere. However, researchers need to more clearly state the scientific problems they are addressing and the metrics they are using for evaluation. The work also needs to be properly placed within the current state of the art. Presenters need to better articulate the primary contributions of the research and how the presented approaches achieve those contributions. The recently hired Ph.D. researchers need to better clarify how their new work at ARL is going beyond their dissertation contributions as students. More broadly, the challenges of robot intelligence specific to the needs of the Army need to be clarified. While the three areas related to mapping, control for communication, and cognition are 55

OCR for page 51
important, they do not provide a perspective of the ARL vision for robotic intelligence for Army applications. In the work on long-duration 3D mapping and navigation, there is a need for justification of why the approach performs better for this problem of long-duration mapping, including identifying metrics against which progress can be gauged. The scalability issues with the approach were correctly identified, and it will be interesting to see how the proposed strategy of forgetting parts of the map helps address this problem. In the area of robust control of mobility and communications, it is important to calibrate the performance of the approach against related methodologies in the fields of mobile ad hoc networking and optimization. There could be some important linkages of this work to the network sciences CTA, and that could present opportunities for leveraging. The design of new cognitive architectures for control can play an important role in robotics. However, it is difficult to understand the benefit of this cognitive approach to the mapping problem. Many working solutions for mapping exist that do not make use of cognitive solutions, and it is not clear how such an approach improves mapping performance. The approach could instead be motivated by a different application domain requiring more high-level cognition, such as a scenario that entails going to the back of a building and watching a door for persons of interest. In the area of HRI interface design, the breadth of experiments is clearly commendable. The experiments, however, are conducted as separate efforts and employ different tasks. This limits the ability to draw meaningful insights from the results. For example, an Android touchscreen interface was compared to an Xbox controller, and separately a speech interface was compared to manual navigation through a GUI. An opportunity exists to place these experiments and results within a larger context and to interpret results across experiments to provide more general design guidance. In HRI research efforts related to understanding automated system actions, the initial experiments provide encouraging results for simple task scenarios involving few factors at a time. There is a need for follow-on experiments that validate the use of visual aids when performing more complex tasks, particularly in a multitasking environment, and on dismounted soldier interfaces. It is important to conduct more basic research that takes advantage of ARL’s unique access to soldiers. As all of the services move toward the inclusion of more robot systems, it is necessary to conduct the basic research that will allow these systems to be effective and efficient members of the team. It was heartening to note the existing collaborations with researchers in cognitive architectures and perception. This model needs to be replicated across other projects at the 6.1 level, where HRI can help guide the development of capabilities that are still very difficult to achieve without a human in the loop (e.g., perception work). Even with increased system capabilities, both in terms of intelligence and perception, there will still be a need for a soldier to interact with the robot systems. The nature of the HRI will change, from remote operation to a supervisory role, and eventually to interaction with the robot as a team member. HRI, however, will remain the key to the effective deployment of robots in the Army and other services. The HRI group would benefit greatly from wider exposure. ARL could consider sponsoring a workshop that would discuss HRI from the soldier’s perspective. In addition to inviting academics and people from the other service laboratories, sponsors of HRI research from other agencies such as the Office of Naval Research, National Science Foundation, Army Research Office, and Defense Advanced Research Projects Agency could bring valuable insight to this effort. On a more general level, much of the work presented was mature; new opportunities for providing input and feedback on projects that were in their inception stage could be beneficial to the overall research enterprise. 56

OCR for page 51
OVERALL TECHNICAL QUALITY OF THE WORK Many of the ARL internal research projects in the autonomous systems enterprise are of very high quality and have benefited from engagement with other research institutions, including partners in the CTAs. For each of the key areas—perception, intelligence and planning, human–robot interaction, and manipulation and mobility—the overall technical quality of the work is high and is being recognized as such by virtue of publication in archival journals and proceedings of recognized conferences and symposia. Also, the recent Research@ARL monograph series on autonomous systems 4 is commendable. For most of the work reviewed, the scientific quality is comparable to that available at other federal research laboratories and at universities here and abroad. The research staff is very well qualified to undertake the research projects and is broadly aware of the state of the art in the field and ongoing research at other institutions. A number of research scientists have been newly recruited, and they promise to contribute to an exciting future for the ARL. Mentorship for these early-career scientists will be of paramount importance for their long-term success at ARL. The laboratory facilities and the infrastructure are state of the art and supportive of the ongoing research activities. The internal research efforts at ARL are being performed against the backdrop of a research roadmap that includes contributions from partners and contractors. It would be useful to clarify this context in order to better understand any gaps that might exist in the research approach. Other specific suggestions to improve the overall research enterprise include the following: • Require researchers to clearly articulate the existing technical challenges in their research and how and why the proposed tools and methods are likely to resolve those challenges. • Require all researchers to identify quantifiable metrics against which progress can be gauged. This would allow for setting meaningful goals and adopting a research agenda that targets nonincremental advances. • As ARL continues to build its research staff, give some attention to bringing in midcareer and senior personnel to mentor the outstanding early-career scientists who have been recruited. • Look for additional ways to increase interaction of its researchers with leaders in industry and academia, given that limitations on travel have restricted this important function. • Focus on developing a mature framework to guide the conception, design, development, and testing of small, unmanned autonomous systems, including definitions of pertinent parameters and their domain (values). • Adopt a systems integration approach as a fundamental research thrust. Existing projects would benefit significantly from such a research thrust. 4 Research@ARL: Autonomous Systems. Available at http://www.arl.army.mil/www/pages/172/docs/ Research@ARL_Autonomous_Systems_July_2013.pdf. Accessed September 20, 2013. 57