automation or autonomy promote effective human-computer interaction [15]. However ..... to the prosecution of a potentially deadly campaign using autonomous ...
Design Challenges for an Autonomous Cooperative of UAV’s Anthony Finn, Kuba Kabacinski, Samuel Picton Drake & Keith Mason Defence Science & Technology Organisation Department of Defence, Australia
Abstract The Defence Science & Technology Organisation (DSTO), which is part of the Australian Department of Defence, is developing a research capability that uses small, inexpensive, autonomous Uninhabited Air Vehicles (UAV’s) to detect, identify, target, track, and electronically engage groundbased targets such as radars. The UAV’s, which act autonomously and cooperatively, use a geographically distributed and heterogenous mix of relatively unsophisticated Electronic Warfare (EW) sensors and other miniaturised payloads networked together to deliver a distributed situational awareness picture that can be shared across the command echelons. The cooperation and networking of these platforms and payloads provides results potentially superior to those of the significantly more expensive, platform-centric systems, but with the added advantage of robustness. This paper reports on the architecture of the DSTO multiUAV cooperative and the status of its development to date. The paper also discusses the future challenges relating to autonomy, supervision, and control that the developers face.
INTRODUCTION A series of strategic analyses of defence capability requirements and priorities have repeatedly concluded that the effectiveness of the Australian Defence Force (ADF) depends primarily upon its combined exploitation of new technologies, doctrine, and geography. Consequently, the future ADF is likely to be designed, as now, to comprise highly trained, well equipped personnel, selected for their resourcefulness and ability to improvise. However, the availability to Defence of skilled human resources is likely to decline over the next fifteen to twenty years in Australia. This means that the size of the future ADF can be expected to remain numerically small and that the contributions of the warfighters to attaining operational objectives must be as high as possible [1]. As a small defence force, the ADF is also likely to continue to depend upon a small number of high value assets to act as force multipliers. For survivability, these platforms have historically depended upon organic sensors and data links to
maintain situational awareness, with Electronic Warfare (EW) self-protection used as a last resort. Unfortunately, increasingly capable air, land, and sea combat systems use a combination of sensors networked together to provide an adversary with the capacity to precisely track and target these high value assets at long ranges. Moreover, as the ADF is a small defence force, attrition has a disproportionately high impact on capability. As a result, the projection of EW in such a way as to have an effect beyond the immediate selfprotection of the platforms is becoming vital to the survival of these high value assets. In order to achieve this projection, the warfighter needs extended reach and access into regions traditionally denied to him. Such systems will inevitably comprise expendable, unmanned systems networked together and to the highervalue, manned assets. Drivers such as cost effectiveness and a move towards single operator, high-workload environments will also ensure that these systems depend upon a number of automated technologies. Overall, therefore, these new autonomous and semi-autonomous systems offer higher operational effectiveness per warfighter with lower casualty levels. In the longer term they also have the potential to reduce the cost of acquisition and operations. The aim of the research described in this paper is to explore the cost-capability trade-offs between larger, more sophisticated, platform-centric situational awareness options and smaller, cheaper, distributed, network-centric ones. The development of this capability requires that a large number of UAV’s be controlled by a small number of people: a concept that implies a significant degree of autonomy. A degree of human decision-making is preferred over complete autonomy, however, because at present humans are better able to issue high-level goals, manage uncertainty, and inject a degree of creativity and flexibility into the system. In addition to the complexities of networking the sensors and automating the platforms, therefore, the fused, value-added situational awareness product must also be disseminated to users across the command echelons in a manner that is meaningful to people. This must be accomplished in real time to allow monitoring and redirection of the UAV’s (as well as for battlespace purposes).
MULTI-UAV COOPERATIVES Distributed situational awareness derived through the use of autonomous and unmanned vehicle systems are likely to enhance early entry operations such as the Suppression of Enemy Air Defence (SEAD). SEAD involves the preparation of the battlefield through the incapacitation of enemy radars, surface-to-air missiles, communications links, etc. It is traditionally considered to be one of the more hazardous military undertakings and it is easy to imagine an air strike platform participating in a SEAD mission protected by a number of UAV’s employing advanced sensors and a combination of hard and soft-kill payloads. The broad requirements of such a system are: that it is persistent, low cost, stealthy, and readily deployable and retrievable; that it can detect, locate, track, identify, characterise, and engage targets autonomously; that it can gather, disseminate, and act on several different types of information; that the individual platform and sensor elements can self-organise; and, that it does not impose significant risk or burden upon the operators. There are clearly a range of platform, mobility, propulsion, and energy issues that need to be addressed for such a system. These are not dealt with in this paper, except to note that the shortcomings and vulnerabilities of larger, slow-moving UAV’s in this context are well known and have been described elsewhere [2] [3]. In summary, a major problem for these larger UAV’s is that their development process parallels that of larger (manned) aircraft, which stresses longer life, a high level of maintainability, a multi-role capability, and high reliability. The resulting systems are expensive with lifecycle costs and logistic complexities approaching those of manned aircraft. To develop the airborne component of their experimental capability DSTO purchased six miniaturised Aerosonde UAV’s (www.aerosonde.com.au). These Affordably Expendable1 UAV’s [3] offer the prospect of a capability that allows exploration of the high value, high risk missions that are beyond the capability (or justifiability) of other systems. There is, of course, no free lunch. Although stealthy, low cost, long endurance, miniature UAV’s and their associated payloads lend themselves to a capability to be placed in harm’s way, the smaller, less expensive, lighter systems are generally significantly less capable than their larger, more strategic counterparts. Moreover, they also carry less capable payloads. However, this may be offset by the increased affordability of the systems, our ability to network the UAV’s and sensors to derive process gain, and our capacity to withstand losses due to conflict or due to malfunction. The requirement for the network of sensors to be mobile and adaptive means that a large number of entities must be 1
The concept of affordable expendability relies upon the notion that the useful life of the capability is a function of its constituent payloads and technologies rather than the physical life of the airframe.
controlled by a relatively small number of people, which in turn requires a high degree of autonomy. This means that each platform must use onboard sensors and processing to capture, represent, and interpret the relevant environmental cues (eg. location, geometry, spectral content, etc), and then autonomously combine and manipulate this information such that the result is a series of control actions that represent both the individual and collective priorities of the systems, subsystems, and payloads within the cooperative. Some of the considerations for such a system are: The number of UAV’s is potentially large Scalability is desirable as UAV’s may leave or join the cooperative or teams Humans need to set goals for and interact with the UAV’s and their payloads The health of the UAV’s, their sub-systems, and their payloads need to be monitored Each individual UAV in a team needs to possess its own complex behaviour Each team within the cooperative should possess its own complex behaviour Humans can be supervisors and/or controllers of UAV’s and/or payloads Supervisors and UAV’s are potentially distributed over a wide geographic area The cooperative should exhibit a highly fluid teamtasking and structure There may be various supervisors (of varying authority) Operations will occur in an adversarial environment Situational awareness events can require quick responses There is a high probability of losing resources The above presents a goal of UAV’s working seamlessly together and with manned vehicles to provide a common, coordinated situational awareness picture. However, its instantiation will require significant scientific advances. One of these is coordinating the actions of the UAV’s that will be carrying a heterogenous set of payloads. Existing multi-robot coordination algorithms are not well-suited to this as most such systems elicit emergent behaviour such that the individual robots follow simple coordination rules rather than teamwork models or goals. These techniques usually break down because the UAV’s cannot explain their actions or role to other members of their team [4]. Moreover, as described above, the concept implies almost complete autonomy whereas we foresee the need for the warfighter to be retained within the decision-making cycle. In addition to the integration of the sensors and platforms, therefore, the information must also be combined, suitably manipulated, and passed to a rear echelon, where it is further integrated with user applications such as geospatial information, track data and imagery, and visualisation and document management tools. This fused, value-added product must then be disseminated to users in near real time to allow the redirection of the UAV cooperative.
Another reason for retaining the human in the decisionmaking cycle is the desire to have a degree of unpredictability in military systems. From a human operator’s perspective, consistency is desirable, but in an adversarial context the capacity to predict exactly what the teams will do may be unhelpful. Consequently, a balance must be struck between consistency, unpredictability, and explanation that still allows the operator to understand and trust the UAV cooperative’s or individual UAV’s action’s at any time [5].
CHALLENGES FOR MULTI-UAV COOPERATIVES Fundamentally, a multi-UAV control system needs to perform three basic tasks: don’t let the aircraft hit the ground; don’t fly the aircraft beyond their limits; and, don’t let the aircraft collide with one another [6]. Once these priorities have been accommodated, higher order tasks (surveillance, reconnaissance, target location, identification, mission coordination, communication, etc) may be undertaken.
As stand-alone actions, the priority tasks are accommodated relatively easily as their goals are both readily decomposable and quantifiable as parameters relating to physical quantities and closed loop control laws (see later). Unfortunately, an effective autonomous multi-UAV cooperative requires the basic airspace tasks to be closely integrated with the higher order tasks and a major impediment to the successful execution of the higher order tasks is the decomposition of the coupled actions that are required to perform an activity such as cooperative surveillance. In addition to this, there is a need for multiple levels of feedback control [6], which in turn depend upon the capacity of individuals and the cooperative to measure and prioritise their performance and actions against a number of metrics (which need to be adequately defined in the first place), and their ability to communicate (in a meaningful fashion) the success of these endeavours, both internally within the group and externally to the human operators. The critical supervisory element for an autonomous cooperative of UAV’s is the feedback mechanism that allows the human operators to understand what, how, and why the system behaves like it does. To survey a region of interest a UAV cooperative must undertake the following tasks: At launch, based on some a priori information about target distribution and priorities allocated by a commander, the mission planning software must generate a series of near-optimal trajectories for each of the UAV’s to follow such that they visit as many regions of opportunity as possible, while simultaneously avoiding as many hazards as possible. The optimisation of these trajectories must be based on (potentially time-varying) cost functions that allow for such things as: the distribution of payloads within the cooperative; the prioritisation of targets; the robustness
of the proposed solution to operational and environmental uncertainties; the individual capabilities of the participating platforms; the benefits that derive from the association of the UAV’s into teams; the communications and sensor scheduling requirements between the platforms to enable this cooperation: and, airspace deconfliction requirements. Once flying their missions, based on a change in the environment observed by one or more sensors onboard the UAV’s, the system must respond by dynamically recalculating trajectories, re-allocating task/team associations, and enabling payload and/or platform actions based on the manipulation and fusion of the new data. Similarly, based on a change in the environment provoked by one or more of the UAV’s payloads or actions (e.g. jamming, UAV’s joining the group), the system must dynamically re-calculate their trajectories, associations, etc. Based on a change in an operator’s priorities or task objectives the system must respond by dynamically recalculating their trajectories, associations, etc. Finally, all of the computational processing and communication must be achieved within the physical and electrical resources of the UAV and in real time.
The main benefit of carrying out these tasks using an autonomous cooperative of UAV’s is the ability of the networked system to process large amounts of information in a relative short period of time to more optimally achieve the high-level goals of target location, ID, and engagement (while simultaneously protecting the users from potentially high risk environments). In order to effectively execute these tasks, however, there are a number of issues [7] - [32]: The potential for information overload This program ultimately depends upon the control of a large number of autonomous or semi-autonomous unmanned vehicles by a relatively small number of human operators. Consequently, the degree of system automation required is fundamentally defined by the relationship between the human resource supplied and the situational awareness task demanded. This means that the UAV’s need to share their individual perceptions of the environment by developing and maintaining a common situational awareness picture. This information must also be filtered, manipulated, and then presented in such a way that a user can quickly assess the status of both the UAV cooperative and the battlespace it observes. If the mental resources required to accomplish this exceed the task demand, system performance remains above the required threshold. In a high workload environment, however, when the demand imposed by the competing tasks exceeds a supervisor’s capacity to process information, performance can be expected to suffer 2 [7].
2
[7] Also predicts a significant drop-off in performance for low workload environments.
To this end, automation should only be introduced where it replaces the “difficult” task responsibilities and presents the residual tasks to operators appropriately. The problem, of course, is identifying the difficult, high priority tasks for such a dynamic decision-making environment and then collecting, processing, storing, and disseminating the information to those who need it, be they on or removed from the battlespace. Selection of the difficult responsibilities is dependent upon a number of factors that include the nature of the task, operational tempo, levels of operator training, and experience. Modelling and simulation tools – in particular hardware in the loop simulations – are a useful aid in the evaluation of the cognitive saturation points of the humans and the overall systems performance of the user-cooperative combination. This also requires the development of a framework and a set of metrics that enables the research results to be evaluated in the form of experiments with data that is quantitative or fiducially referenced 3. The degree of autonomy required Much research has been conducted into what levels of automation or autonomy promote effective human-computer interaction [15]. However, there is very limited research into the support requirements for teams of “swarming” autonomous vehicles [13]. For rigid tasks that require limited flexibility in decision-making, and with a low probability of system failure, higher levels of autonomy often provide the best solutions [16]. However, in systems that must deal with dynamic environments that have many external and changing constraints, higher levels of automation are not usually successful because of their inability to reliably expedite the decision-making process in the face of uncertainty or unforeseen problems [17]. It is predominantly this that drives the need to include a human in the supervisory loop. One of the main issues for the interaction of the vehicles and the human supervisor, therefore, is the impact of the human decision-making process on the system performance. If the automation is not highly reliable, the UAV cooperative may perform poorer than one with no automated assistance [18]. Moreover, several studies have demonstrated the human tendency to rely upon computer-based recommendations, even though there may be contradictory (and usually correct) information readily available. This is known as decision bias and is discussed in another section. There are two components of automation involved in the supervision of a cooperative of UAV’s: one described by Sheridan & Verplank 4 [19] and one described by Cummings5
3
There are few metrics, benchmarks, or standardised tests against which the performance of autonomous or unmanned systems can be measured or compared. Currently, most results are in the form of demonstrations or comparisons to human equivalence. 4 Sheridan & Verplank describe ten levels of computer-human interaction varying from “The computer offers no assistance, the human must make all decisions and actions” to “The computer decides everything and acts autonomously, ignoring the human”
[11]. Consequently, when the UAV cooperative is tasked with locating and engaging a target, and the UAV’s do not have any capacity to collaborate, the levels of supervisorUAV interaction can still vary from one to ten against the Sheridan-Verplank scale (i.e. the automation can vary from “the human makes all the decisions” to “the UAV’s make all the decisions”). Alternatively, in the case where there is full intra-UAV collaboration, the human-UAV interactions must exist only at the higher levels of the Sheridan-Verplank scale and the UAV cooperative would determine the best candidate to engage the target. This duality in the levels of automation presents a difficult supervisory control problem. In single vehicle problems, there are ten discrete levels of autonomy which can allow direct comparisons of the system’s overall performance to be made against one another. However, when there are networks of vehicles the problem space becomes significantly larger and more complex. Consequently, when designing a support system that allows the humans to interact with vehicle cooperatives it is important to assess the impact of the levels of human-UAV automation, the effects of various levels of collaboration between the UAV’s, and the indirect influences of interaction between the automation schema. A requirement for adaptive automation In a dynamic system, resources must be balanced against time constraints and the relative importance of the tasks that must be undertaken. In a distributed situational awareness system, such as that described here, the limited cognitive processing capabilities of the supervisors is one such resource. There are two issues that must addressed, both related to task intervention: information expected from the cooperative has not arrived and a decision must be made (i.e. the user must act on incomplete information); and, the user simply has multiple dynamic tasks that vary in priority and must be balanced against one another. In regard to the first of these issues, recent studies [21] into high-workload, NCW-related environments have revealed that more research is required. The initial results indicate that when the users are supplied with decision aids that (accurately) predict future periods of high workload they fixate on attempts to globally optimise an uncertain future schedule to the detriment of solving specific, local problems. In regard to the second issue, a number of studies involving target confirmation requests [22] [20] have concluded that the cognitive load on operator is likely to become saturated when controlling between 13 and 16 UAV’s. These studies all used UAV teams/systems and had limited collaborative control 5
Cummings describes four levels of intra-vehicle collaboration/autonomy varying from “The vehicles do not communicate with one another and follow original tasking unless a human identifies a new task” to “The vehicles are in full collaborative communication, and individual vehicle tasking changes according to a predetermined algorithm: there is no human intervention”
and/or decision-aiding for the operators. However, those studies carried out on systems that have team performance monitoring and prompts to the user to intervene when trouble is detected [14] have indicated that a user can effectively supervise much larger teams of UAV’s 6. The scheme simply identifies situations where human input might be needed and then explicitly transfers responsibility for decision-making to the human. Typically, these (human) decisions require projections into the future or global judgements that are unlikely to be considered by the reactive and/or localised nature of the decentralised processes. Because a human may not be able to respond to the prompts in a timely manner, the value of the decision may be lessened. Consequently, mathematical models of the transfer of control that capture the increasing appropriateness of terminating a long-running plan need to be employed [23] [24]. The Need for Decentralised Decision-Making The fundamental building block of good decision-making is a high degree of situational awareness, which may be defined as having three levels: the perception of the elements in the environment; the comprehension of the current situation; and, the projection of the future status [9]. For distributed decision-making a high degree of shared situational awareness is required. In an NCW environment the devices that deliver shared situational awareness include spoken and non-verbal communications, visual and audio shared displays, and a shared environment [25]. Unfortunately, the bulk of these delivery mechanisms are not viable for an autonomous cooperative of UAV’s and the sensed data must be pre-processed to convert it to a common reference frame, fused with state predictions based on historical observations, transmitted through communications interfaces, assimilated with other sensed data that have passed through a similar process to that described here, and then represented visually for interpretation and use by the cooperative’s supervisor. The benefits of this highly complex process are scalability, robustness, fault-tolerance, and flexibility. Although significant advances are required in this area to instantiate this element of the program, there is a significant body of work relating to the decentralised data fusion and sensor scheduling requirements available upon which to base our endeavours [8] [26] [27]. Flight path deconfliction & collision avoidance Flight-path deconfliction and collision avoidance for multiple UAV’s imply similar route re-planning requirements separated by their time of action. Airspace deconfliction is a medium range task that attempts to avoid a collision while still allowing the UAV to remain within some predetermined navigation corridor, maintain time-on-target, conserve fuel, etc. Collision avoidance is a last minute, emergency manoeuvre aimed solely at preventing aircraft loss. The latter does not take mission completion into its decisions [28]. 6
These experiments involved simulations with teams of eighty UAV’s.
Initially, at least, multi-UAV operations require the aircraft to be designated pre-defined flight paths, regardless of whether or not they have the ability to cooperate with one another. Irrespective of whether these pre-defined flight paths are generated using a route-planning algorithm or manually by an operator, the resultant trajectories must conform to acceptable levels of airspace deconfliction in terms of the temporal and spatial separation between the aircraft. If the UAV’s are networked and can coordinate their efforts, then one or more of them may dynamically and continuously adapt their flight paths (e.g. in response to a target detection) to increase the effectiveness of their overall search strategy and/or capacity to prosecute the targets. Consequently, even if only one UAV needs to deviate from its pre-planned trajectory (autonomously or under the control of an operator), the rest of the cooperative must also have the capacity to dynamically adapt their trajectories safely. Alternatively (or additionally) constraints may be placed upon the degree to which the manoeuvring UAV’s are allowed to adapt their trajectories. The collision avoidance algorithm must also be able to accommodate “blunders”, where one aircraft deviates from its intended path for unforeseen reasons. In this case, other UAV’s must then manoeuvre to avoid collision or maintain adequate separation. Generalised solutions to coordinated control and assignment problems for UAV’s are non-trivial, particularly when there are cooperation constraints imposed (e.g. communications ranges, min/max airspeed velocities, collision avoidance, etc). Moreover, these generalised solutions do not usually lend themselves to an extension of simple two-UAV control and assignment problems [30]. Regardless, the environment must be monitored and the appropriate state information collected and disseminated within the cooperative. These states then provide an estimate of the current situation (i.e. UAV position and velocity). A dynamic trajectory model is then required to predict the future states of the UAV’s so that these states can then be combined to evaluate the likelihood of conflict. The outcome is the continuous (e.g. every 30 sec) calculation of a guaranteed period (e.g. up to 5 min) of conflict-free trajectory for each UAV while it carries out its higher order mission. Cooperative mission & trajectory planning Ideally, given an area of interest, an automatic route planning algorithm will calculate search patterns for the group of UAV’s, optimised under constraints such as: maximise the probability of detection; minimise the time to detection, minimise the number of UAV’s required; maximise the robustness of the search to aircraft loss; and/or minimise the amount of network traffic required. The search algorithms must also account for areas of threat and terrain obscuration. Based on a priori knowledge of targets within an area of interest an operator must designate the location, dimensions, and orientation of an area of interest (AOI) and whether or not it is known to contain targets. Thereafter, based on the payload configuration, flight trajectories, target locations,
priorities, etc for each of the UAV’s involved in the search are calculated such that the probability of detecting the targets must be maximised. Dynamic cooperative control Once the UAV’s have commenced flying their pre-planned trajectories, they must then autonomously and dynamically respond to the detection of targets and threats. These responses may be stimulated by either their own onboard sensors or those onboard other UAV’s in the cooperative. Alternatively, the UAV’s may need to respond to a change in priority imposed externally by the supervisors. In order to do this, each UAV must have goals and priorities assigned, which in turn requires a set of metrics that enables them (individually and collectively) to evaluate situations and events. These metrics enable the UAV’s executive controllers (colloquially know as agents7) to autonomously choose between competing goals, to assign resources to the task, and generate priorities that maximise pay-off and minimise cost. Ideally, both predicted and observed situations are evaluated so that resources can be allocated ahead of time and “standoffs” and conflicts can be avoided. At this time, however, the agents that have been trialled in the real UAV’s are capable of only “one step ahead” computations. The formation and dissolution of teams In addition to the UAV’s responding dynamically as individuals or collectively within the cooperative to sensed opportunities and threats, in order to maximise the system’s effectiveness the cooperative also needs the capacity to form teams. The motivation for team-formation is to improve the probability of attaining specific goals – i.e. the detection or geolocation of an emitter. In order to keep the supervisory workload to a minimum, however, the team forming process needs to be self-organising, such that the formation of a team is the result of forces acting within the cooperative and between the member UAV’s, as opposed to being imposed externally by an operator. Two other attractive properties of self-organisation are that the formation can potentially perform self-repair and that it can respond appropriately when unusual events occur. Health & usage monitoring One of the primary consequences of the separation between the operator and the aircraft is that the operator is deprived of a range of sensory cues that are available to the pilot of a manned aircraft. Consequently, rather than receiving the sensory input directly from either the vehicle or the environment in which the vehicle is operating the UAV operator receives only that sensory information provided by onboard sensors via a data link. The sensory cues that are lost include visual information, kinaesthetic/vestibular input, and As used here, “agents” are self-contained code units that interact with one another and their environment. The techniques used to perform the priority tasks must be verifiably safe and reliable: isolation of the higher order tasks into small, self-contained code units also allows for easier verification
sound [9]. For instance, actuator malfunction may be signalled to the pilot of a manned aircraft via visual, auditory, and haptic feedback. In contrast, for a UAV this failure may be indicated solely by perturbations of the camera image. To address this “sensory deprivation” a considerable amount of data must be relayed from sensors and systems onboard the UAV to the operators at the GCS. This data must then be processed and presented in such a way as to simultaneously minimise the operator’s workload in regard to the need to monitor it and maximise his capacity to interpret and understand it. Alternatively, onboard HUMS (health & usage monitoring systems) must autonomously process, interpret, and deliver meaningful information about the status of the UAV, its sensors, and its sub-systems to the users. Monitoring systems performance In addition to performing all of the above we also need to evaluate the performance of the systems, their constituent components, and the human interactions with them. In other words, we need to evaluate the quality of the system’s process in addition to its end results. We are likely to achieve this by providing an insight into competencies that effective teams possess: knowledge, skills, attitudes, leadership, and decision-making [15] [28] [31] [32]:
Task Distribution – Is the human taking control where appropriate and letting the automation handle its duties when appropriate; and vice versa? Strategies – To solve a problem, was an appropriate strategy selected? How appropriate was the strategy and how well was it executed? Decisions – How quickly are they made? How ‘correct’ are they? Directions – How quickly can they be submitted? Did the system do what the human expected it to do? Adaptability – How well does the cooperative handle unforeseen events or errors? Alternatives – How many valid alternative strategies were derived and presented to the user? How varied were the alternatives? Communication – What is the quality of communication within and between teams? What is the complexity of the information flows? How much interpretation is necessary for the human to apply information from the system? How much work is needed for the human to direct the automation? How much human resolution of ambiguities and uncertainties is required? Support – Do the team members exhibit good feedback and backup behaviour? Situational Awareness – Is there agreement between the human and automation’s perception of the situation? To what extent does the human always accept the operational picture without question, or never accept it?
7
There are a number of other future challenges that require further detailed investigation. These include:
Mitigating the potential for decision biases Humans are very good at issuing high-level goals, managing uncertainty, and injecting a degree of creativity and flexibility into systems. Unfortunately, they are also prone to decision biases that are heavily influenced by experience, the framing of cues, and the presentation of information [20].
Supervising and monitoring the operators The main issue here is how to provide diagnostic and feedback support to the supervisors of the UAV’s supervisors who may themselves be distributed over a wide geographic area [25].
Trust and reliability Trust in automation is a dynamic that changes with time. Moreover, users frequently trust malfunctioning equipment and/or mistrust equipment operating correctly. A great deal of research into this area is required [13].
Legality & Accountability In addition to the significant technical challenges, there are a number of issues of legal accountability pertaining to the prosecution of a potentially deadly campaign using autonomous systems, particularly given the issues of trust and reliability referred to above [11].
DSTO’S EXPERIMENTAL MULTI-UAV SYSTEM In conjunction with Australian industry and academia, DSTO has developed an experimental multi-UAV system that can be commanded locally or remotely by any of a number of users using high level command objectives. Figure 1 shows a block diagram of the current implementation of the system. The colour coding provides a guide to the EW (blue), EO (purple), and synthetic environment (green) components.
Figure 1: Functional Block Diagram of DSTO's Multi-UAV Distributed Situational Awareness & Targeting System.
pan-tilt-zoom electro-optic (EO) payload, high resolution, still-image EO camera payloads, multi-aspect, fixed orientation and gimballed infra red (IR) payloads, an RF repeater payload (used as a controlled radar cross-section in conjunction with the electronic attack (EA) payloads), acoustic detection payloads (for vehicle and small-arms fire detection and/or countermeasures), a tactical data radio relay payload, and a high resolution scanning laser radar (LADAR). Suitably combined, the information from these sensors has the potential to provide a rich situational awareness picture from the air that includes ID and intent, EO and IR imagery, through-foliage target detection and location, and high resolution, 3-D target information. The DSTO Distributed Situational Awareness and Targeting (DSAT) system is geared around the principles of Network Centric Warfare (NCW), deriving its effectiveness from a combination of modularity, distributed sensing, sensor-target geometry, and “system” processing gain. The system exploits the inherent range advantages enjoyed by EW payloads to cross-cue those with shorter ranges (e.g. EO/IR payloads). It also has the capacity to receive information from third-party sensors and to autonomously task EA payloads. The current implementation of the multi-UAV ES system, known as NERDS (Networked EW & Radar Detection System) uses a centralised-decentralised architecture that communicates through the DEWSAT server8. In other words, although many of the computation processes take place in a single location – and sometimes even on a single PC – the architecture is fully representative of a completely decentralised (and scalable) system. For instance, the frontend signal conditioning and processing, the Emitter Library Management (ELM), geolocation, tracking, and fusion elements of the system run on single PC’s. However, the PC’s run separate and parallel computation processes that can (and have) run on separate PC’s connected only via a LAN or RF modems that emulate the communications processes between the UAV’s. This allows for integration of the communications package into the UAV’s as a miniaturised processing capability or a slightly larger UAV system becomes available9. The architecture also lends itself to the processing taking place on a manned platform or a larger “mother ship” UAV. The architecture of the EW component of the DSAT system is shown in Figure 2. The colour-coding is as follows: green represents the sensors and other low-level data sources, blue the low-level data processing and dissemination, gold the low-level tasking & control, purple, the cross-system fusion, system tasking, and dissemination control, and white the high-level automation, user-tasking and visualisation. The 8
A number of payloads have been developed that include: communications and radar electronic surveillance (ES) payloads, communications and radar jammers, a stabilised
DEWSAT [38] is a distributed server which implements service protocols to enable distributed command & control and situational awareness. It is the effectively the backbone of the DSAT system. 9 The signal conditioning and processing elements are currently migrating to miniature processors onboard the aircraft.
SDWs
ELM Classify emitters Locate emitters
PDWs
ES O
bje
ct
SLIDE Collect and disseminate PDWs and SDWs
3rd Party Sensors PDWs
Ta
Intel, ESObject, ESSensor Platform, Task, EAeffector
sk, ES Se ns
or ,
Inte l
Falcon View Create tasks
Task Agent Teaming Assign assets Create new tasks (sub-tasks) Monitor tasks
PDWs
UAV payloads
GUAVA Collect PDWs Tune ES sensors
Task
Steer Agent Generate flight plans
Tune
Task, AssetPool
DEWSAT Server Task, Platform ESObject, Intel
EAEffector,EOObject EOSensor, ESObject
ESSensor, Intel Platform, Task
waypoints
Aerosonde GPS
Figure 2: Functional Block Diagram of NERDS The general tasking sequence is as follows: High level instructions are issued by the operator to search an area of interest, for which there may or may not, be a priori information about potential targets; Based on the a priori information, the payloads onboard each of the UAV’s, and the specific tasking instruction, the DSAT system autonomously selects a sub-set of UAV’s from the cooperative to execute the task(s); Using the broadband IFM receivers all of the emitters within view are coarsely classified in terms of their spectral content. If a priori information is available about the location, this is associated with the spectral information autonomously; Priorities are allocated to each of the emitters. There are “rules” that allow the emitters to be autonomously prioritised, but generally this is input by a user; Based on the priority and location of the targets and the geometry of the UAV’s available for this task, each of the UAV’s within this sub-set is then autonomously assigned a task/target using one of a number of teaming algorithms [36]; By combining the information observed and interpreted by the UAV’s organic ES sensor, and that presented to it by its colleagues in the team, all of the emitters in view can be coarsely located. The shared information is then also used to calculate the individual trajectories of the platforms, which now instantiate a responsive, cooperative surveillance schema that derives from the decomposition of six lower level “steering” behaviours [35]: o Approach target – steer the UAV towards a target to provide better signal-to-noise ratio (SNR) for the ES receivers (this also improves geolocation geometry); o Flee target – steer the UAV away from the target to avoid its detection; o Coalesce – steer the UAV to maintain communications with its neighbours;
Separate on distance – steer the UAV to maintain a minimum safe distance between itself and the other UAV’s to allow for collision avoidance; o Separate on angle – steer the UAV to improve the geometry for geolocation; o Avoid “no fly” zones – steer the UAV’s to avoid forbidden airspace; The resultant steering vector is the vector sum of the spatially continuous distribution generated by the UAVtarget geometry and other influences. This scheme ensures the system continues to operate in the presence of uncertain, incomplete, time-varying, and some-times incorrect information; As the UAV’s fly their trajectories they continue to share information such that the accuracy with which they geolocate the target radars progressively improves; At present no airspace deconfliction algorithms are applied to the search templates generated by the algorithms described above. These are incorporated into the steering computations that respond to the sensor processing of the UAV’s payloads. The individual UAV’s within a team are re-allocated tasks (based on [36]) when the geolocation accuracy achieves a user-specified threshold, when it is unable to converge further, or when a user decides to re-prioritise the tasking of the UAV’s. When the distributed ES system has accurately geolocated the target(s) a UAV carrying a stabilised PanTilt-Zoom (PTZ) EO payload can be used to overly, monitor, and “prosecute” the relevant targets. o
connections between the layers are through a combination of analogue, serial, UDP, TCP/IP, and SOAP interfaces.
Unfortunately, the fidelity of the native image observed by miniaturised PTZ payload is limited by the quality of the optical and electronic components of the system. For instance the images can suffer from blurring effects and are also inherently limited in their resolution by the detection array and the optics used to gather the image. Specifically, the density of cells in the detection array fundamentally limits the resolution of the gathered images. Details in the image plain smaller than the size of a cell are averaged out during the image capture process. Consequently, we apply image enhancement techniques to overcome the limitations of these imaging systems through the fusion of a set of low-resolution images (related to each other through random translations and rotations in the image plane) in order to create a higher resolution image of the original scene. This provides targetquality imagery/data. The user interface to the UAV’s is enabled locally via FalconView, which has been extensively modified. Through the use of an additional toolbar, the user has access to all of the normal functionality of FalconView, but additionally can control individual UAV’s, payloads, and/or teams of UAV’s by assigning them areas of interest (search areas), individual waypoints or targets, exclusion zones, and other spatially meaningful instructions. The user can also control the cooperative remotely (i.e. anywhere with access to the www)
via web-based interfaces through DEWSAT (to FalconView) and through Personal Data Assistants (PDA’s) running standard web-browsers. Information sensed by the UAV’s is also simultaneously integrated and disseminated to these user interfaces to enable command decisions to be made and/or modified in real time. At present, should differing command decisions be made at different locations, there is no algorithm implemented for autonomously resolving the conflict. Information from the above system has also been integrated with DSTO’s Force Level EW Synthetic Environment (FLEWSE), such that a combination of real and synthetic UAV’s can interact with one another in a single environment. As this test bed also has the capacity to run the same code as that flown on the UAV’s, it also allows pre-flight testing of key elements of software as well as mission rehearsal. It is also a very cost effective way of “flying” large numbers of UAV’s to experimentally evaluate some of the complex (coupled) automation and human supervisory control issues associated with the development, operation, and optimisation of the multi-UAV based DSAT system.
RECENT DSTO TRIALS In a series of trials at the Woomera Test Range, the following has been demonstrated using the above: In October 2004, a heterogeneous mix of miniaturised ES payloads (one Instantaneous Frequency Measurement receiver (IFM) and two Super Heterodyne (SH) receivers), distributed across three Aerosonde UAV’s were used to monitor the spectrum from 2-18GHz. The IFM onboard one UAV detected a (stationary) radar target and cross-cued the SH receivers onboard two other UAV’s to tune to and monitor the same target radar emissions. The pulse descriptor words (PDW’s) describing the target radar emissions and computed by the ES payloads were then correlated and used to coarsely geolocate the radar using scan ranging and Time Difference of Arrival (TDOA) techniques [28]. Once the target was coarsely located, the UAV’s then autonomously, dynamically, and continuously adapted their flight trajectories to progressively improve the accuracy with which they were able to co-operatively geolocate the rapier battery. The precise geolocation of the target radar in the RF domain was then used to cross-cue another UAV carrying the stabilised PTZ EO payload, which was then used to “prosecute” the target from the air. In March 2006, five Aerosonde UAV’s (four fitted with ES sensors and one with an EA payload) were flown simultaneously. A further two stationary ES payloads were incorporated into the cooperative. A larger target set of emitters were also deployed. A teaming algorithm [36], was then used to dynamically and autonomously allocate the six ES sensors (four on UAV’s and two stationary – “on sticks”) into two teams of three sensors that were then able to separately geo-locate the prioritised radar targets. The ES
sensors onboard the UAV’s detecting and locating the radar emitters were then used to cross-cue a UAV fitted with an Electronic Attack (EA) payload, which successfully engaged the targets. The situational awareness picture derived from the cooperative was injected into a simulation of the Suite of Integrated RF Countermeasures (SIRFC) obtained from the US under Project Arrangement 10 (PA-10). SIRFC is a stateof-the-art electronic combat system fitted to helicopters. It provides threat awareness to a pilot on the basis of fusing onboard and off-board ES data and also controls countermeasure systems. Information from the real UAV’s operating at Woomera was also injected into FLEWSE (in real time) such that the real UAV’s and a number of simulated UAV’s were able to interact with pone another in the EW scenario.
CONCLUSION This paper presents a rationale for autonomous, multi-UAV based distributed situational awareness and targeting systems and discusses a range of issues relevant to supervisory control that the developers must face in the future. It also describes DSTO’s experimental distributed situational awareness & targeting system, which uses small, inexpensive, UAV’s to detect, identify, target, track, and electronically engage ground-based targets such as radars. The paper also provides a status report on some recent trials carried out by DSTO. ACKNOWLEDGEMENT The authors are grateful to their colleagues, Kim Brown, Jean-Pierre Gibard, Leszek Swierkowski, Hatem Hmam, Andrew Bailey, Damian Hall, Luke Marsh, Don Gossink, Greg Calbert, Roger Flint, Po-Young Wong, and Jeremy Fazakerley who are jointly responsible for developing the autonomous multi-UAV system described here. Their many contributions to this paper are also gratefully acknowledged.
REFERENCES [1] S. Allison, Options for DSTO Research in Automation of the Battlespace, DSTO Discussion Paper, May 2003 [2] A. Finn, Automation of the Battlespace Initiative, CDSAC Discussion Paper, 2004 [3] Tactical EW Division, NRL, Affordably Expendable UAV’s, Paper prepared for US Secretary of Navy, April 1999 [4] Nourbakhsh et al, Human-Robot Teaming for Search & rescue, Pervasive Computing, IEEE Computer Society, January-March, 2005 [5] A. Finn, K. Brown, & A. Lindsay, Miniature UAV’s & Future Electronic Warfare, Proc. Land Warfare Conference, Brisbane, October 2002 [6] S.R. Dixon & C.D. Wickens, Control of Multiple UAV’s - A Workload Analysis, Proceedings 12th International Symposium Aviation Psychology, Dayton, OH (2003)
[7] R.M. Yerkes & J.D. Dodson, Relationship of Strength to Stimulus to Rapidity of habit Formation, J. Comparative Neurology & Psychology, Vol. 18, 1908 [8] D. Lambert, Ubiquitous Command and Control, Proceedings Information, Decision and Control Conference, Adelaide, South Australia,1999 [9] A. Miller, Modelling Human Workload Limitations on Multiple UAV Control, Proc. 48th Annual Meeting of Human Factors & Ergonomics Society, Santa Monica, 2004 [10] J.S. McCarley, C.D. Wickens, Human Factors Concerns in UAV Flight, University of Illinois Human Factors Division Technical Report AHFD-05-05/FAA-05-1 [11] M.L. Cummings, Supervisory Control Challenges in NCW, University Virginia, Doctoral Dissertation, 2003 [12] J.V. Nickerson & R.R. Reilly, A Model for Investigating the Effects of Machine Autonomy on Human Behaviour, Proc. 37th Int. Conf. on Sys Sci, Hawaii, April 2004 [13] M.J. Ashleigh & J. Nadhakumar, Trust and Technologies: Implications for IT Supported Work Practices, Proc. ECIS Conf, June 6-8, Poland, 2002 [14] M. Lewis, J. Polvichai, K. Sycara, & P. Scerri, Scaling Up Human Control for Large UAV Teams, Proc. Human Factors & Ergonomics Soc. 48th Annual Meeting, 2004 [15] T.B. Sheridan, Telerobotics, Automation, & Human Supervisory Control, Cambridge, MA, MIT Press, 1992 [16] M.R. Endsley & D.B. Kaber, Level of Automation Effects on Performance, Situational Awareness, and Workload in a Dynamic Control Task, J. Ergonomics, Vol. 42(3), pp462-492, 1999 [17] C.D. Wickens, Automation in Air Traffic Control: The Human Performance Issues, Automation Technology & Human research trends (Eds: M. Scerbo & M. Mouloua), Lawrence Erbaum Assoc., 1999 [18] S.R. Dixon & C.D. Wickens, Imperfect Automation in UAV Flight Control, Institute of Aviation, AFHD Technical Report, AFHD-03-17/MAAD-03-2, 2003 [19] T.B. Sheridan & W. Verplank, Human & Computer Control of Undersea Teleoperators, Cambridge, MA, Man-Machine Systems Laboratory, Dept Mechanical Engineering, MIT, 1978 [20] M.L. Cummings, Human Supervisory Control of Swarming Networks, Proc. 2nd Conference Autonomous Intelligent Networked Systems, Arlington, VA, 2004 [21] P.M. Mitchell, M.L. Cummings, T.B. Sheridan, Mitigation of Human Supervisory Control Wait Times through Automation Strategies, Technical Report, MIT Humans & Automation Laboratory, Massachusetts Institute of Technology, MA, 2005 [22] M.R. Endsley & D.B. Kaber, Level of Automation Effects on Performance, Situation Awareness, and Workload in a Dynamic Environment, Ergonomics, Vol. 42(3), 1999 [23] P. Scerri, D. Pynadath, M.Tambe, Twowards Adjustable Autonomy in the Real World, J. Artificial Intelligence Vol. 17, pp 171-228, 2002
[24] P. Scerri, D. Pynadath, M.Tambe, Adjustable Autonomy in the Context of Coordination, Proc. AIAA 3rd Unmanned Unlimited Technical Conference, 2004 [25] M.R. Endsley & W.M. Jone, A Model of Inter- & IntraTeam Situational Awareness, New trends in Cooperative Activities (editor M.R. Endsley), Human Factors & Ergonomics Society, 2001 [26] J.H. Sutcliffe & J.M.A. Waldie, Real-time decentralised data fusion: a framework for future tactical operations, Proc Land Warfare Conference, p.249-257, Brisbane 2002 [27] H.F. Durrant-Whyte, S. Clark, M.W.M.G. Dissanayake, P. Newman and M. Csorba, An experimental and theoretical investigation into simultaneous localisation and map-building (SLAM), Proc 6th Sym. on Experimental Robotics, Sydney, 1999 [28] R. Teo et al, Project CAREFUL – Scenarios & Measures of Effectiveness, Joint DSO/DSTO Research Report, February 2005 [29] S.C. Allen, C.R. Jones, & S. Guerlain, The Effect of Command & Control Team Structure on Ability to Quickly & Accurately Retarget UAV’s, Proc. 2nd Conference on Autonomous Intelligent Networked Systems, Arlington, VA, 2004 [30] J. K. Kuchar, L. C. Yang, A Review of Conflict Detection and Resolution Modeling Methods, IEEE Transactions on Intelligent Transportation Systems, Vol.1, No. 4, 2000 [31] P.J. Hinds, T.L. Roberts, H. Jones, Whose Job Is It Anyway? A Study of Human-Robot Interaction in a Collaborative Task, J. Human Computer Interaction, Vol. 19, 2004 [32] J.S. McCarley & C.D. Wickens, Human Factors Implications of UAV’s in the National Airspace, Aviation Human factors Division, Institute of Aviation, Technical Report AHFD-05-05/FAA-05-01, April, 2005 [33] H. Hmam, Scan-Based Emitter Passive Localization, IEEE Trans. on AES, Vol. 43, No 1, 2007 [34] S. Drake et al, Agent Development for 2006, DSTO Internal Publication, 2006 [35] S. Drake, K. Brown, J. Fazackerley, & A. Finn, Autonomous Control of Multiple UAV’s for the Passive Location of Radars, ISSNIP Conference, Melbourne, December, 2005 [36] L. Marsh, D. Gossink, S. Drake, G. Calbert, UAV Team Formation for Emitter Geolocation, submitted to Proc. IDC 2007, Adelaide, February, 2007 [37] A. Finn, et al, Automation of the Battlespace: An Integrated Air-Ground, Multi-Vehicle, Multi-Sensor, Autonomous UVS Demonstration, AUVSI Conference, Anaheim, 2004 [38] K. Brown et al, A Demonstration of Distributed Sensing & Situational Awareness, Land Warfare Conf., 2006
