3
Autonomy Technology: Capabilities and Potential

Autonomous vehicles (AVs) have demonstrated that they can significantly increase the operational capabilities of modern armed forces, and it is evident that they will become an even more important element of warfighting capability in the future. This chapter discusses the state of the art of autonomous systems, examines some promising autonomy technology that will be available in the near future, and identifies some shortfalls in autonomy capability that need to be alleviated. The chapter goes on to explore the level of autonomy as a design choice and autonomy technologies.

TODAY’S AUTONOMOUS VEHICLE SYSTEMS

Types of Systems

There are three types of autonomous vehicle systems: scripted, supervised, and intelligent. Scripted autonomous systems use a preplanned script with embedded physical models to accomplish the intended mission objective. Examples of these systems include smart bombs and guided weapons. Such systems can be generally described as “point, fire, and forget” systems that have no human interaction after they are deployed.

Supervised autonomous systems automate some or all of the functions of planning, sensing, monitoring, and networking to carry out the activities associated with an autonomous vehicle, while using the cognitive abilities of human operators via a communications link to make decisions, perceive the meaning of



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 45
Autonomous Vehicles in Support of Naval Operations 3 Autonomy Technology: Capabilities and Potential Autonomous vehicles (AVs) have demonstrated that they can significantly increase the operational capabilities of modern armed forces, and it is evident that they will become an even more important element of warfighting capability in the future. This chapter discusses the state of the art of autonomous systems, examines some promising autonomy technology that will be available in the near future, and identifies some shortfalls in autonomy capability that need to be alleviated. The chapter goes on to explore the level of autonomy as a design choice and autonomy technologies. TODAY’S AUTONOMOUS VEHICLE SYSTEMS Types of Systems There are three types of autonomous vehicle systems: scripted, supervised, and intelligent. Scripted autonomous systems use a preplanned script with embedded physical models to accomplish the intended mission objective. Examples of these systems include smart bombs and guided weapons. Such systems can be generally described as “point, fire, and forget” systems that have no human interaction after they are deployed. Supervised autonomous systems automate some or all of the functions of planning, sensing, monitoring, and networking to carry out the activities associated with an autonomous vehicle, while using the cognitive abilities of human operators via a communications link to make decisions, perceive the meaning of

OCR for page 45
Autonomous Vehicles in Support of Naval Operations sensor data, diagnose problems, and collaborate with other systems. Most conventional autonomous vehicles and their controlling elements form an autonomous system that fall into this category. Intelligent autonomous systems use intelligent autonomy technology to embed attributes of human intelligence in the software of autonomous vehicles and their controlling elements. This intelligent autonomy software does the following: (1) it makes decisions, given a set of (generally automated) planned options; (2) it perceives and interprets the meaning of sensed information; (3) it diagnoses vehicle, system, or mission-level problems detected through monitoring; and (4) it collaborates with other systems using communications networks and protocols. This major section discusses technologies relating to supervised and intelligent autonomous systems. The systems and technology associated with such systems generally reside in the Mission Management System or Command and Control System elements of an the autonomous system (see Figure 3.1), while the actions that implement higher-level decisions are done today (generally autonomously) by the Vehicle Management System (VMS) (e.g., by autopilots). Following is a descriptive list of the various systems that comprise the elements of an AV system. FIGURE 3.1 The elements of an autonomous vehicle system. NOTE: C2, command and control; C4ISR, command, control, communications, computers, intelligence, surveillance, and reconnaissance; MCG&I, mapping, charting, geodesy, and imagery; ECM, electronic countermeasures; FLT CNTL, flight control; SA, situation awareness.

OCR for page 45
Autonomous Vehicles in Support of Naval Operations Planning and Decision. Planning and decision is the process of developing a sequence of actions capable of achieving AV mission goals or activity goals, given the state of the world. The Planning and Decision System dynamically plans and commands functions within the VMS to carry out mission activities, given situation-awareness information from the Sensing and Perception System and self-awareness information from the Monitoring and Diagnosis System. A plan diagnosis assesses the need to replan on the basis of situational changes derived from updated information. Planning and decision systems often use human-machine collaboration to complete their tasks. Sensing and Perception. The Sensing and Perception System collects, fuses, and interprets sensor data from local sensors and from the Networking and Collaboration System, which receives data from external sources. This information is used to develop a mission-relevant picture or digital map representation of the current mission situation for use by the Planning and Decision System. The digital map, which is dynamically updated, contains information on the location of the AV with respect to all known threats, targets, terrain, obstacles, and friendly forces. Sensing and perception systems often use human-machine collaboration to complete their tasks. Monitoring and Diagnosis. The Monitoring and Diagnosis System collects, fuses, and interprets sensor information relating to the health of the AV. Its responsibilities include the fault detection and isolation (FDI) of system, subsystem, or component failures. FDI helps prevent loss of the AV resulting from system failures and increases the probability of mission success if vehicle systems can be reconfigured during a mission using redundant capability. This system may also include sensors to monitor health trends in key subsystems in order to enable preventive maintenance and prognostication of future failures. Networking and Collaboration. The Networking and Collaboration System manages the use of data links, frequencies, and information content for purposes of collaboration. Collaboration involves the sharing of information with other autonomous or manned vehicles operating as a team or with other vehicles operating in the same space. The types of information shared are, for example, navigation state for collision avoidance, pop-up threat locations, new target locations or targets of opportunity, and vehicle mission plans or plan fragments required to support the collaboration. Human-System Interface. The Human-System Interface System is an extremely important element of an autonomous system. Even in highly autonomous systems, humans are required to provide high-level objectives, set rules of engagement, supply operational constraints, and support launch-and-recovery operations. Humans are also needed by autonomous systems to help interpret sensor information, monitor systems and diagnose problems, coordinate mission time lines, manage consumables and other resources, and authorize the use of weapons or other mission activities.

OCR for page 45
Autonomous Vehicles in Support of Naval Operations Other Autonomous Behaviors. Some VMS functions contain autonomous modes or behaviors that can be commanded and controlled by the Planning and Decision System. A common example is an autopilot function of the Guidance, Navigation, and Control System, which may have multiple modes depending on the flight phase, flight conditions, or operating environment. The State of the Art Contemporary autonomous systems employ a wide range of autonomy technology, depending on the vehicle domain (i.e., air, ground, sea) and the operating requirements of the system. The following subsections present a brief summary of the current state of the art for the autonomy capability areas developed in the preceding section, “Types of Systems.” Planning and Decision The general problem of planning and decision has been addressed in operations research and artificial intelligence for more than 30 years, with the research addressing increasingly complex formulations of the planning problem. Path planning or route planning is commonly available today in all domains. Autonomous mission planning, which involves the development of plans to achieve mission goals, is primarily accomplished through automated tools that are defined premission and subsequently executed. The Navy’s Portable Flight Planning System (PFPS) for aircraft is an example of a planning system in use today. The PFPS and the developmental Joint Mission Planning System (JMPS) are excellent premission flight-planning systems with large databases of information to support high-fidelity flight planning; however, both lack the ability to rapidly accommodate evolving mission events through dynamic planning. The modification of mission plans owing to the occurrence of unanticipated events is heavily dependent on “humans in the loop” for all autonomous vehicle domains. Dynamic mission planning that enables autonomous mission replanning to take into account unanticipated events is not common today, although capabilities on unmanned undersea vehicles (UUVs) have advanced the state of the art in this area. Dynamic mission-level planning is also a current thrust in the Office of Naval Research’s (ONR’s) Maritime Reconnaissance Demonstration (MRD) Program and its Intelligent Autonomy Program (e.g., the Risk-Aware, Mixed-Initiative Dynamic Replanning Program). Some collaborative multivehicle planning development, at a low level of autonomy, has also been done in the past for unmanned aerial vehicles (UAVs) at ONR in the Uninhabited Combat Air Vehicle (UCAV) Demonstrations Program and at the Air Force Research Laboratory (AFRL) in the Cooperative Manned/ Unmanned Systems Program. Both programs used a single ground station to

OCR for page 45
Autonomous Vehicles in Support of Naval Operations control a team of UAVs that shared Global Positioning System (GPS) navigation solutions for route deconfliction. Finally, the National Aeronautics and Space Administration’s (NASA’s) Remote Agent Experiment was executed for several days onboard the NASA Deep Space One mission,1 representing a significant demonstration of autonomy in space operations. This mission emphasized planning and decision capabilities to maintain the spacecraft in a desired internal state by planning time lines of activities, sequencing lower-level steps together to achieve higher-level goals, and executing plans in a reliable fashion. The system made use of probabilistic models of the subsystem hardware to detect and diagnose failures and replan the mission activities. Temporal planners, such as the Remote Agent Planner, can take hours to generate plans of large size unless hand-coded heuristics are provided, but alternatives are under development to improve searches for feasible time bounds of mission activities when generating mission time lines. Sensing and Perception Sensing and perception technology in today’s fielded systems is primarily used for AV navigation and avoidance of terrain hazards. Most AVs employ GPS-aided inertial navigation systems, although UUVs also employ Doppler velocity logs or other velocity correction sensors to aid the inertial system for navigation. Terrain sensing—using sonar for UUV bottom following and unmanned ground vehicle (UGV) behaviors such as wall following or road following—is also in use today. Cruise missiles employ terrain-matching and scene-matching technology that may have application for some UAV missions. Obstacle-detection technologies have also been a research focus over the past decade, with emphasis on AV operations in complex terrain. This capability is particularly important for off-road UGV operations, littoral UUV operations, urban environment UAV operations, and undercanopy UAV applications. Obstacle-detection systems use a variety of sensors, including electro-optic cameras (stereo and mono), infrared cameras, ultrawideband radars, sonars, and light detection and ranging (LIDAR). The ONR Maritime Reconnaissance Demonstration Program is using bathymetry maps and forward-looking sonar to perform obstacle avoidance. The Defense Advanced Research Projects Agency (DARPA)/Army Demonstration III Program employed LIDAR and stereo cameras to build a three-dimensional map of the vehicle’s immediate surroundings, which was then used to plan local paths that move toward a goal while avoiding the obstacles. Autonomous systems that detect, classify, and identify targets or threats are limited primarily to the UUV domain, although manned aircraft also include 1   For further information, see the Web site <http://nmp.jpl.nasa.gov/ds1/>. Last accessed on April 5,

OCR for page 45
Autonomous Vehicles in Support of Naval Operations technologies to support the pilot that could be utilized for UAVs. The creation of situation-awareness maps is also rare today, except in UUVs used for mapping the location of underwater mines, which was done in the mid-1990s in the DARPA Autonomous Minehunting and Mapping Technologies Program and is a part of the Remote Environmental Monitoring Unit System (REMUS), Remote Minehunting System (RMS), and Long-range Mine Reconnaissance System (LMRS). The ONR Maritime Reconnaissance Demonstration Program (part of the Autonomous Operations Future Naval Capability (FNC)) is using a situation-awareness sensor suite, including communications intelligence (COMINT), electronic intelligence (ELINT), and video to detect, map, and avoid surface threats. A Virginia-class submarine (VSSN) provides mission command and control for the UUV. The MRD UUV transmits the threat type, location, and bearing to the VSSN, which provides the new threat information to update the battlegroup’s common operational picture. The VSSN also provides target-identification objectives to the MRD UUV for searching out and verifying surface targets. This capability was demonstrated in April 2003 during Fleet Battle Experiment Kilo. Much work has been done and is still ongoing in the area of automatic target-recognition and threat-detection systems. Many techniques have been explored for a variety of sensors, but most methods are limited in their capability owing to unfavorable lighting conditions, weather, and viewing geometry, or obscurations such as foliage or terrain. Still, it is likely that some of this research will be used to field automatic target-cueing systems in the near term. These systems will not likely be fully autonomous, but will help either to increase operations tempo or to reduce operator workload. Monitoring and Diagnosis As described above, monitoring and diagnosis systems are used to detect and isolate failures within AV subsystems. The monitoring and diagnosis systems in use today primarily employ built-in test equipment to sense the malfunctioning of subsystems and equipment. This information is generally used for diagnostics and maintenance support, but is also infrequently used to support the reconfiguration of the autonomous system or the replanning of the mission, particularly in UUVs. System reconfiguration and mission replanning typically require redundant systems to be available onboard the AV. Some UUVs today also make use of triplex or quad-redundant, fault-tolerant computers that choose among input and output signals to detect and isolate failures. This technology, more common in manned systems, is infrequently used today for autonomous vehicles. DARPA’s Autonomous Minehunting and Mapping Technologies Program was an example of the use of quad-redundant, fault-tolerant computing in a UUV. Analytical redundancy—which makes use of mathematical models of hardware subsystems to provide estimates of the expected sensor measurements or

OCR for page 45
Autonomous Vehicles in Support of Naval Operations vehicle responses for failure detection and isolation—is employed in manned systems, but is infrequently used in autonomous vehicles today. Networking and Collaboration Most of today’s AVs do not directly or autonomously collaborate with other manned or unmanned vehicles. Those that do primarily exchange navigation state to permit collision avoidance with other vehicles and often do so through ground control stations with human intervention. Collaboration among vehicles is largely accomplished by the operators controlling the mission. Research on networking and collaboration for AVs has increased in recent years, with programs such as DARPA’s Mobile Autonomous Robot Software (MARS)2 and Software for Distributed Robotics (SDR).3 These programs are researching soft computing, initiative learning, coordinated control, and networking and communications autonomy technology to enable future collaborative robot capabilities. LEVELS OF AUTONOMY In order to classify systems for purposes of comparison, it is useful to identify the level of autonomy (LOA) that systems exhibit. Defining LOA in a simple, useable form has proven to be a difficult task. As yet, no single scale expressing LOAs has been found acceptable across the broad range of users. Intuitively, it seems that the mix of human and machine capabilities to be found in any particular system (or vehicle) implementation could be appropriately characterized by position along a linear axis with manual operation at one end and fully autonomous operation at the other. The many such attempts to define simple LOAs in this fashion have resulted in scales with differing numbers and definitions of the intermediate levels. These scales are summarized below, together with an expanded view of LOA as recommended by the committee. Autonomy Scales Defined by the Department of Defense One level-of-autonomy scale, created by the DARPA/U.S. Air Force (USAF)/Boeing X-45 program team, represents a rather high-level, broad-brush view of autonomy, with only four levels. This scale is presented in Box 3.1. 2   For additional information, see the Web site <http://www.darpa.mil/ipto/programs/mars/vision.htm>. Last accessed on April 5, 2004. 3   For additional information, see the Web site <http://www.darpa.mil/ipto/programs/sdr/vision.htm>. Last accessed on April 5, 2004.

OCR for page 45
Autonomous Vehicles in Support of Naval Operations BOX 3.1 Levels of Autonomy as Defined by the Uninhabited Combat Air Vehicle Program Level 1 (Manual Operation) The human operator directs and controls all mission functions. The vehicle still flies autonomously. Level 2 (Management by Consent) The system automatically recommends actions for selected functions. The system prompts the operator at key points for information or decisions. Today’s autonomous vehicles operate at this level. Level 3 (Management by Exception) The system automatically executes mission-related functions when response times are too short for operator intervention. The operator is alerted to function progress. The operator may override or alter parameters and cancel or redirect actions within defined time lines. Exceptions are brought to the operator’s attention for decisions. Level 4 (Fully Autonomous) The system automatically executes mission-related functions when response times are too short for operator intervention. The operator is alerted to function progress. Another, more detailed LOA scale, with 10 levels, was created by the Army for the Future Combat System (FCS) Program. That scale is shown in Table 3.1. Still other LOA scales similar to these have been created by other programs in connection with developing autonomy technology or autonomous vehicles. These include the Air Force’s autonomous control levels, which are defined for the observe-orient-decide-act (OODA) loop.4 The OODA loop defines different LOAs for each of the four primitive elements of closed-loop autonomy, namely—observe, orient, decide, and act. The intermediate levels of one scale often seem to be unrelated to those of another, so a one-to-one correspondence between the levels defined by different scales is difficult to establish. The source of this confusion lies in the one-dimensional nature of most attempted definitions of LOAs, as well as in the 4   For additional information, see the Web site <http://www.adtdl.army.mil/cgi-bin/atdl.dll/fm/6-0/appa.htm>. Last accessed on April 5, 2004.

OCR for page 45
Autonomous Vehicles in Support of Naval Operations differing focus of each of the groups defining the LOAs. The application of autonomy concepts and technology to a system is inherently a complex issue, with several degrees of freedom that must be addressed. Thus, it is impossible to characterize the implemented degree of autonomy completely with a single number. An Expanded View of Level of Autonomy The main expectation for Navy and Marine Corps autonomous vehicles is that they be able to carry out mission goals reliably, effectively, and affordably with an appropriate level of independence from human involvement. However, in practice it is difficult to assign a single level of autonomy to any AV. This is largely because AVs and their controlling systems are designed to perform complex missions made up of many activities, each of which may be implemented with a different level of autonomy. This fact implies that the notion of complexity must also be considered when assigning an LOA to an AV. This section proposes a new view of level of autonomy, which is hereafter called the level of mission autonomy. As described below, mission autonomy is made up of two degrees of freedom—mission complexity and degree of autonomy. “Mission complexity” captures the number of functional mission capabilities inherent in any given system or the number of different mission activities that can be implemented by the system, independent of whether they are accomplished autonomously or not. “Degree of autonomy” captures the amount of autonomy used to implement any specific mission activity or functional capability. Mission complexity, the first degree of freedom, is not to be confused with system complexity, which increases as the number and variety of system elements (e.g., vehicles, operators, processors, data links, sensors, databases, power bases, and so on) become greater and as the level of predictability of the system decreases. System complexity results, in part, from the selection of mission autonomy requirements. To further elaborate on mission complexity, it is useful to view it in the context of an autonomous vehicle mission. A mission is a hierarchical collection of mission activities that are sequenced to accomplish mission goals. High-level activities (i.e., mission phases such as launch, ingress, operations, egress, and recovery) are broken down into subordinate activities, which are themselves further decomposed into primitive activities. Each mission activity can be accomplished by a different mix of human and/or machine collaboration. The human involvement in the mission can be categorized in terms of control and authorization, coordination, and intelligence, as the examples in Box 3.2 suggest. The number of mission activity levels (e.g., high, medium, low), the number of mission activities within each level, and the degree of human-equivalent functionality (e.g., intelligence) required for each are design choices that, once made, define the complexity of the AV itself. Mission complexity is then characterized by the number of functional mission capabilities that can be performed by the

OCR for page 45
Autonomous Vehicles in Support of Naval Operations TABLE 3.1 Levels of Autonomy in the Army Scale for the Future Combat System Level Level Description Observation Perception and Situation Awareness Decision-Making Ability Capability Example 1 Remote control Driving sensors None Remote operator steering commands Basic teleoperation 2 Remote control with vehicle state knowledge Local pose Reporting of basic health and state of vehicle Remote operator steering commands, using vehicle state knowledge Teleoperation with operator knowledge of vehicle pose situation awareness 3 External preplanned mission World model database—basic perception Autonomous Navigation System (ANS)-commanded steering based on externally planned path Basic path following, with operator help Close path following intelligent teleoperation 4 Knowledge of local and planned path environment Perception sensor suite Local plan/replan—world model correlation with local perception Robust leader-follower with operator help Remote path following—convoying 5 Hazard avoidance or negotiation Local perception correlated with world model database Path planning based on hazard estimation Basic open and rolling semiautonomous navigation, with significant operator intervention Basic open and rolling terrain 6 Object detection, recognition, avoidance or negotiation Local perception and world model database Planning and negotiation of complex terrain and objects Open, rolling terrain with obstacle negotiation, limited mobility speed, with some operator help Robust, open, rolling terrain with obstacle negotiation

OCR for page 45
Autonomous Vehicles in Support of Naval Operations 7 Fusion of local sensors and data Local sensor fusion Robust planning and negotiation of complex terrain, environmental conditions, hazards, and objects Complex terrain with obstacle negotiation, limited mobility speed, and some operator help Basic complex terrain 8 Cooperative operations Data fusion of similar data among cooperative vehicles (such as UAVs) Advanced decisions based on shared data from other similar vehicles Robust, complex terrain with full mobility and speed. Autonomous coordinated group accomplishments of ANS goals with supervision Robust, coordinated ANS operations in complex terrain 9 Collaborative operations Fusion of ANS and reconnaissance, surveillance, and target acquisition (RSTA) information among operational-force UGVs Collaborative reasoning, planning, and execution Accomplishment of mission objectives through collaborative planning and execution, with operator oversight Autonomous mission accomplishment with differing individual goals and little supervision 10 Full autonomy Data fusion from all participating battlefield assets Total independence to plan and implement to meet defined objectives Accomplishment of mission objectives through collaborative planning and execution, with operator oversight Fully autonomous mission accomplishment with no supervision SOURCE: LTC Warren O’Donell, USA, Office of the Assistant Secretary of the Navy (Acquisition, Logistics, and Technology), “Future Combat Systems Review,” presentation to the committee, April 25, 2003.

OCR for page 45
Autonomous Vehicles in Support of Naval Operations Simultaneous Localization and Mapping Navigation in GPS-denied environments has received considerable attention in recent years in order to improve the navigation of UGVs in most operating environments, that of UAVs in urban environments or under tree canopy, and that of UUVs in littoral waters (in Chapter 5, see the section entitled “Naval Operational Needs and Technology Issues,” for UUVs). Sophisticated processing means are becoming popular for combining the functions of navigation and mapping to improve the quality of both. SLAM is a technique by which terrain objects or topography are entered into a map at the same time that the position and orientation of the vehicle is being estimated in those same map coordinates. A crucial effect of this technique is that when a piece of terrain (e.g., a feature or object) is seen again after the vehicle has moved significantly, the system performs a correlation between the old observations and the new, giving simultaneously a tremendous improvement in the map accuracy and in the vehicle navigation state. Such techniques can give highly accurate estimates of vehicle position and terrain topography. Furthermore, cooperative execution of such algorithms by multiple vehicles sharing a common data structure can quickly produce high-quality maps and localizations for all of the vehicles. This technique has been applied in relatively structured environments (e.g., inside buildings or tunnels), where features are noncomplex and easily recognized, using LIDAR, sonar, and vision sensor systems. This technology is less mature for operations in unstructured environments where features or map objects are of various shapes and sizes. Most of the SLAM work to date has used commercial off-the-self sensors and focused on algorithm and software development. But in most cases the sensors involved are not in a form suitable for fielding. Thus, there is a significant gap in sensor development, particularly for intelligent autonomy for small UGVs. The Army Research Laboratory’s Collaborative Technology Alliances Program is funding sensors germane to vehicles the size of the FCS Multifunction Utility Logistics Equipment (MULE) vehicle or larger.10 Threat Detection and Identification As autonomous vehicles become more accepted, they will be called on to operate in more threat-dense environments. Real-time capability for threat detection and identification will be required for AV operations in these environments. Today, manned aircraft, surface ships, and submarines make use of threat radars, electro-optical (EO) and infrared (IR) sensing, and COMINT signal processing to detect and identify adversary threats and threat types. Many of these technologies 10   For additional information, see the Web site <http://www.arl.army.mil/alliances/Default.htm>. Last accessed on May 18, 2004.

OCR for page 45
Autonomous Vehicles in Support of Naval Operations are transferable to AVs to enable operations in threat-dense environments. In order to operate on an AV, these systems will need to be augmented by planners for threat-response tactics that take the place of the pilots or operators to implement one of various strategies in response to a threat. Also, in many cases the sensors used on large, manned vehicles will be too big for AVs, or the ability for autonomous threat detection with high probability of detection and low false alarm rate is not very mature. Thus, more work is needed in this area. Analytical Redundancy and Failure-Detection Filtering Conventional approaches to monitoring and diagnosis of vehicle systems include the use of hardware redundancy for failure detection and isolation using input-output voting schemes, midvalue selection, and built-in testing. These methods by their nature can substantially increase the weight of the Vehicle Management System and do not by themselves help determine the lost functionality within a subsystem or the mode of the system owing to the failure. The latter is critical for a dynamic planning system to be able to determine the right course of action following a failure. Analytical redundancy, which makes use of mathematical models of hardware subsystems to provide estimates of the expected sensor measurements or vehicle responses, does not require redundant hardware and can be used to determine the lost functionality within the affected subsystem. Analytical redundancy provides estimates of the expected sensor measurements or vehicle responses through estimation of theoretical approaches developed beginning in the 1940s and 1950s (e.g., the Wiener filter and the Kalman-Bucy filter). Failure detection and isolation using analytical redundancy employ estimation of theoretical technologies such as hypothesis testing, maximum-likelihood detection, generalized likelihood ratio tests, and robust estimation, to detect and isolate system failures. These methods use linear filters to generate residuals between a model of the system and the measurements being received from onboard sensors. A failure in the dynamic system can be detected as a change in one or more of the plant parameters, or input signals. These faults can correspond to failed actuators or sensors or to failures that cannot be assigned to any system components (e.g., a UUV getting caught in a net). In detection filter design, the filter gain is chosen so that the residual vector has a different fixed direction for each hypothesized component failure. Hypothesis tests describe the expected response of the system to the no-failure case and to selected candidate failures. Ratios of probabilities of the various failures to the no-failure response are computed and compared to a threshold to detect and isolate failures. The generalized likelihood ratio test is a statistical test that looks for a change in the statistical properties of the filter to declare a failure of a specific type. Robust estimation approaches modify the filter gains to accommo-

OCR for page 45
Autonomous Vehicles in Support of Naval Operations date uncertainties in the mathematical description of the subsystem processes being used. These methods have been developed and tested for UUVs and are in use today in aircraft-engine health-monitoring systems, commercial-airline diagnostics and prognostication systems, and the guidance, navigation, and control systems of spacecraft and military aircraft. Supervised Learning and Adaptation/Learning Technology Learning and adaptation technologies have applicability for autonomous vehicle control, mission planning, failure diagnosis, sensing and perception, and collaboration. These technologies have matured over the past two decades to the point of being a useful component technology to improve mission effectiveness for specific mission activities or to improve vehicle survivability for specific critical-failure scenarios. However, this technology has not matured to the extent that it should be viewed as a panacea for the accommodation of unanticipated events for all mission activities. There are three primary categories of learning and adaptation technology: (1) model approximation, (2) supervised learning and adaptation, and (3) reinforcement learning. The technologies within the first and second categories are mature enough today to be used on a limited basis for specific AV functions if the overall mission effectiveness and vehicle survivability will truly benefit from the expanded capability. Technologies within the third category are not mature enough to be used in AVs today. Model approximation (category 1) makes use of connectionist (learning) networks of radial basis functions, sigmoidal functions, or Gaussian functions to represent complex physical processes that are otherwise difficult to model. Model-referenced adaptive control systems make use of this technology to expand the operating space for vehicle control systems and reduce modeling complexity. Learning-based model approximation has been used to model such things as the nonlinear flight dynamics of aircraft for flight control, aircraft jet-engine combustion for failure detection and isolation, helicopter gearbox models for failure detection, and chemical propagation for the detection and tracking of underwater plumes. Learning-based model approximation has also been used to generate models within planning systems, for state estimators, or for analytical failure detection and isolation. These techniques are heavily supported by simulation data to provide the initial network training, and subsequently they are supported by experiential data collected during the AV’s operations. The technique of supervised learning and adaptation uses a learning system in order to select the best (or a good) action to be implemented, given the current state of the system. The learning is said to be supervised since the selection of a good action uses a network trained through human supervision or simulation. The

OCR for page 45
Autonomous Vehicles in Support of Naval Operations network is trained by computing a value function. The value function is a complex mapping that represents the benefit to be derived by the implementation of each possible action for all possible system states. It can be a mathematical function (e.g., a weighted combination of the system states and possibly previous actions) or the subjective opinion of “goodness,” as determined by a human supervisor. The value function represents the benefit to be derived by implementation of all possible actions. Once the learning system has been trained via this supervision, the system has the ability to generate a good action given an arbitrary system state. Supervised learning systems have been applied to such things as AV controls, mission activity planning, and fault detection and isolation. The technique of reinforcement learning and adaptation is the most difficult and by far the least mature at this stage of development. Reinforcement learning systems are systems capable of learning without access to an a priori provided value function. In this case, the system must learn the value function “on-the-fly,” which requires that trial actions be explored for the inputs that currently exist, and then be quickly evaluated for “goodness.” Many techniques have been developed for this purpose, including Q-Learning and neuro-dynamic programming, but each requires substantial computational resources or processing delays to implement the existing algorithms. Human-Machine Collaborative Decision Making Most autonomous vehicles for the foreseeable future will continue to operate under mixed-initiative control, in which decision making is shared by humans and automated systems. UUVs may be an exception to this rule, owing to the difficulty of communications in the underwater environment. For there to be a force-multiplier effect in the use of AVs, such decision making must involve a single human operator controlling several vehicles. Remote control of every vehicle by a single operator becomes impossible. This level of operator control (or conversely, level of autonomy) is a system design choice, as was previously pointed out. The desired level of human interaction to perform the functions described in Box 3.2 must be selected for each mission activity in the mission activity hierarchy for a particular system of AVs. As the number of vehicles to be controlled increases, so too does the required complexity of the human-machine interaction. The operator must know when to, and then be able to, take more control over mission activities at any time and for any level of the mission activity hierarchy, when required. Similarly, the automated systems must be better able to assess their ability to achieve the desired goals presented by the operator and then request help when needed. This variable or adjustable autonomy will likely be required to enable the Navy’s vision of the future. Technologies available to implement mixed-initiative control today are fairly limited and primarily point solutions to specific portions of the autonomous

OCR for page 45
Autonomous Vehicles in Support of Naval Operations systems. For example, the planning and decision frameworks discussed in the section above entitled “Dynamic Real-Time Mission Planning and Replanning” provide a rudimentary (first) capability for the operator to interact with the system during any mission activity and at any level of the mission activity hierarchy. These interactions can be for the purpose of mission planning, plan execution monitoring, plan problem diagnosis, and authorization of planned activities. Although these frameworks do not preclude the use of variable levels of autonomy for mission activities, they do not presently support this capability either. Similarly, systems that are used to generate situation awareness (e.g., threat detection and response) typically implement a fixed, human-machine interaction protocol. Much work remains in order to develop a system architecture for autonomous systems and the methods that support mixed-initiative control with variable levels of autonomy for planning and decision, sensing and perception, and monitoring and diagnosis. Key Shortfalls in Autonomy Capability Despite the autonomy capabilities that can now be leveraged from the DOD’s autonomy technology portfolio or that are currently being developed via ONR’s Autonomous Operations FNC, much remains to be done if the Navy’s future vision is to be fully realized. The focus of future Naval Services’ investments and the pace of autonomy technology development must be carefully mapped, with cognizance of work being done across the DOD, including work by the Army, the Air Force, and the Defense Advanced Research Projects Agency (DARPA). Table 3.2 lists the top two or three general shortfalls in autonomy capability that need to be remedied in order to enable the operational capabilities described by the DOD’s vision expressed in the 2001 Quadrennial Defense Review11 and in the Navy’s Sea Power 21.12 These shortfalls represent areas in which more intensive, Navy- or Marine Corps-specific development focus may provide the greatest value in enabling new operational capabilities for the Naval Services. For each shortfall in capability, the table lists the level of technology development recommended by the committee, possible future programs (transition targets) that would benefit from the development, a description of the capability needed, and some items to be considered as part of the technology development. Implicit in the recommended level of technology development is the current level of technology maturity that could be built upon to create the new operational capability. 11   Donald H. Rumsfeld, Secretary of Defense. 2001. Quadrennial Defense Review Report, U.S. Government Printing Office, Washington, D.C., September 30. Available online at <http://www.defenselink.mil/pubs/qdr2001.pdf>. Accessed on May 13, 2005. 12   ADM Vern Clark, USN. 2002. “Sea Power 21: Projecting Decisive Joint Capabilities,” U.S. Naval Institute Proceedings, Vol. 128, No. 10, pp. 32-41.

OCR for page 45
Autonomous Vehicles in Support of Naval Operations TABLE 3.2 Key Autonomy Capabilities Shortfalls, by Technology Area, with the Level of Technology Development Recommended by the Committee and Future Programs That Would Benefit Technology Area Shortfalls in Autonomy Capability Recommended S&T Levela Of Benefit to Possible Future Programs Description of Needed Capability Key Considerations Planning and decision Dynamic mission planning for teams 6.2/6.3 LCS Develop a capability for dynamic planning of high-level mission activities involving small teams of vehicles (manned or unmanned). Includes planning for all phases, e.g., launch, ingress, operations, egress, and recovery. Vertical integration of team mission planning with C2, including method of team control (e.g., through master vehicle in master-slave arrangement or through each member of team in peer-to-peer arrangement). How will integrated system deal with targeting? UCAV Mine interdiction warfare systems   Threat-response tactics planning 6.1 UCAV Develop a capability for real-time threat-response-tactics planning, which decides among options involving avoidance of threat, defense against threat with countermeasures, evasion of threat through maneuvering, or attack of threat with available weapons. Explore benefit of dynamic concurrent threat-response planning (concurrent with nominal mission planning) versus using reactive preprogrammed tactics. Sensing and perception Human-machine collaborative threat and/or target identification and classification 6.3 UCAV Integrate currently available algorithms with appropriate sensor modalities to demonstrate automatic target cueing capability. Operationalize existing, but immature, technology by adding human interface to allow human-machine collaboration. Consider focusing on currently available EO, IR, SAR sensor technology. Multi-reconfigurable unmanned undersea vehicle (MRUUV)

OCR for page 45
Autonomous Vehicles in Support of Naval Operations   Human-machine collaborative exploitation of sensor information 6.1/6.2 FORCEnet Develop human-machine collaborative decision-making tools that have increased levels of autonomy for target-of-opportunity detection, classification and/or identification, exploitation in context, verification, and prioritization. Integrate these autonomy tools with system. Consider the distribution of autonomy between autonomous vehicles collecting information, operator control stations, and intelligence exploitation centers.   Sensor development for intelligent autonomy for small vehicles. Perception for autonomous navigation. Monitoring and diagnosis Mission- and/or system-level problem detection, diagnosis, and reconfiguration 6.1/6.2 UCAV Develop a capability to use FMEA, with system component coverage and failure rates to detect and diagnose current and emerging problems in autonomous systems and then to assess the impact of the problem on mission plans. Includes systems of multiple vehicles and communication back to home base(s). Consider multiple levels of autonomy through human-machine collaboration. MRUUV Networking and collaboration Secure, assured networking for multivehicle collaboration 6.1/6.2 LCS Develop a capability to autonomously manage the network of a small team of vehicles collaboratively planning and generating situation awareness. Consider missions in which all vehicles are operating in open terrain and missions in which one or more vehicles are operating in complex environments (e.g., urban environments, underwater, under canopy). UCAV Mine interdiction warfare systems

OCR for page 45
Autonomous Vehicles in Support of Naval Operations Technology Area Shortfalls in Autonomy Capability Recommended S&T Levela Of Benefit to Possible Future Programs Description of Needed Capability Key Considerations Learning and adaptation Real-time learning for adaptation to unanticipated events 6.1 UCAV Develop a capability to learn to adapt “on-the-fly” to unanticipated events. Includes events that were not anticipated but that might occur prior to the mission and for which the value function for several competing responses to the event needs to be learned quickly. Solution may be application-specific. For example, problems of failure reconfiguration may require approaches different from those for problems involving learning and responding to adversary tactics. MRUUV Submarine track and trail Human-system interface Natural user interfaces (e.g., natural language, gestures, symbology) 6.1 General Develop the capability for an autonomous system to understand natural language or gestures of military operators or controllers. Very difficult problem to be solved generally. Consider focusing on specific high-value Navy or Marine Corps needs such as UAV deck operations, manned-unmanned aircraft operations, UGV control via gestures or hand signals, or launch-and-recovery operations.   Variable initiative control 6.2 UCAV Develop the capability for human operators to exert temporal variations of control over missions and activities during mission operations. Multiple levels of autonomy depending on operator workload and vehicle and/or mission state. MRUUV Other multimission capable AVs aScience and technology (S&T) levels: 6.1, basic research; 6.2, applied research; 6.3, advanced technology development. NOTE: A list of acronyms is provided in Appendix D.

OCR for page 45
Autonomous Vehicles in Support of Naval Operations CONCLUSIONS AND RECOMMENDATIONS Autonomous Vehicle Concepts and Developments As discussed above, the Office of Naval Research’s Autonomous Operations Future Naval Capability has initiated a four-pronged autonomy technology development effort. This effort, in concert with the DOD’s autonomy technology portfolio and ongoing DOD programs, provides a pipeline of maturing technologies that can be used to create, in the near term, new Navy and Marine Corps autonomous vehicle capabilities. Some examples include the following: For UAVs and UGVs, the adoption and adaptation of the dynamic real-time mission-planning technology used in UUVs and on spacecraft; The adoption of avionics architectures from spacecraft and manned systems to permit the migration of mission management autonomy software onboard autonomous vehicles; The adaptation of a dynamic real-time mission-level planning module, such as that developed under DARPA Mixed Initiative Control of Automa-Teams or the ongoing DARPA Jaguar Programs, with existing flight-planning systems such as the Navy’s Portable Flight Planning System or the Joint Mission Planning System; The automation of existing manned aircraft threat-detection and -response capabilities for use in autonomous vehicles of all types; The adaptation of existing automatic target-recognition technology to operationalize semiautonomous versions of the technology using human collaboration; and The use of analytical redundancy and the built-in test and diagnostics capabilities in subsystem equipment to provide enhanced system reliability. Autonomous Vehicle Technologies The focus of future Naval Services investments and the pace of autonomy technology development need to be carefully mapped, with cognizance of work being done across the DOD, including that of the Army, Air Force, and DARPA. Table 3.2 lists some of the shortfalls in autonomy capability that need to be remedied in order to achieve the Navy’s future vision—in these areas the committee believes that development focused on Navy-unique capabilities is required to raise the maturity of the technology to moderate levels. The committee believes that investments are needed in those technologies that improve the following: The ability for AVs to operate in threat-dense environments, The ability for human operators and/or intelligence analysts to collaborate with computers to interpret and exploit AV sensor data,

OCR for page 45
Autonomous Vehicles in Support of Naval Operations The ability of AVs to network and collaborate with other autonomous and manned vehicles, The ability of AVs to detect and diagnose mission- and system-level problems and to reconfigure in order to accommodate them, The ability of AVs to perform multiple missions, and The ability of UAVs and UUVs to perform autonomous shipboard launch-and-recovery operations. Incorporate Level of Mission Autonomy as Autonomous Vehicle Design Trade-off System designers of autonomous vehicles often neglect the potential operational benefits to be derived by employing level of mission autonomy as a design choice in up-front trade-off studies, instead electing to focus on trade-offs relating to vehicle performance characteristics (e.g., speed, range, endurance, stealth) and subsystem capability (e.g., sensing and communications). This approach constrains the level of autonomy that can be implemented later in the development and prevents designs that might provide greater operational benefit in terms of impacting mission effectiveness, vehicle survivability, and system afford-ability. Early-stage AV design trade-offs can include the vertical integration of the AV system with its command-and-control system for the end-to-end operations to be performed by the system, including allocation and assignment, mission tasking (e.g., intelligence, surveillance, and reconnaissance; strike; logistics), collection, exploitation, and dissemination. Including the level of mission autonomy as a design choice enables several additional benefits to be derived, such as these: Prioritized, targeted technology development investments for Navy and Marine Corps autonomous vehicle needs based on determining those technologies that will have the greatest benefit; Reduced system complexity achieved through an increase in onboard mission autonomy; Improved autonomous vehicle mission effectiveness and survivability resulting from shorter planning and decision-making cycles; faster assimilation and interpretation of sensor information; faster detection, isolation, and assessment of system problems; shared mission objectives among collaborators; and expanded use of offboard sensor information; and Reduced total cost of autonomous vehicle ownership resulting from reduced operator support for planning, decision, and collaboration; reduced operator support for sensor interpretation and exploitation; reduced operator support for monitoring and problem diagnosis; reduced maintenance labor for trouble-shooting and prognostication; higher system reliability and reduced probability

OCR for page 45
Autonomous Vehicles in Support of Naval Operations of loss of vehicle; and shared use of distributed resources (e.g., sensors, weapons, and so on). Autonomous Technology Recommendations Recommendation: The Assistant Secretary of the Navy for Research, Development, and Acquisition (ASN(RD&A)) and the Chief of Naval Research (CNR) should direct the Navy and Marine Corps Systems Commands, the Office of Naval Research (ONR), and the Marine Corps Warfighting Laboratory (MCWL) to partner with the operational community and monitor the concepts and development of critical autonomous vehicle-related technologies considered essential to the accomplishment of future naval missions. The progress of these developments should be tracked year to year. Specifically: Pursue New Autonomy Concepts and Technology Developments. The ASN (RD&A) should direct appropriate agencies in the Navy and Marine Corps to formulate and maintain a list of the most promising moderately to highly mature autonomy technologies (Technology Readiness Level: TRL > 4) that can enable, critical near-term autonomous vehicle capabilities. Plans to pursue further development of these capabilities should be developed and funded, and progress should be tracked year to year to ensure the proper pace of development. The ONR should develop autonomous vehicle research and development (R&D) needs and a technology roadmap to achieve the goals defined by the various vision documents of the Naval Services. ONR should leverage the current operational experience and the recommended increase in future operational experience with autonomous vehicles in order to define R&D needs to address specific, high-value operational needs. Recommendation: The Assistant Secretary of the Navy for Research, Development, and Acquisition (ASN(RD&A)) should mandate that level of mission autonomy be included as a required up-front design trade-off in all unmanned vehicle system development contracts. Specifically: Incorporate Level of Mission Autonomy as an Autonomous Vehicle Design Trade-off. The ASN(RD&A) should direct appropriate agencies in the Navy and Marine Corps to exploit level of mission autonomy as a degree of freedom for impacting concepts of operations, mission effectiveness, vehicle survivability, and system affordability by including a level of mission autonomy as a design choice in the early-stage system trade-off studies. The architecture of all new autonomous vehicles should be such that increasing levels of autonomy can be implemented in the field by modular replacement and/or software upgrade.