AAAI Technical Report FS-12-04 Human Control of Bioinspired Swarms Controlling Swarms of Unmanned Vehicles through User-Centered Commands Gilles Coppin François Legras Télécom Bretagne – Institut Mines-Télécom CNRS UMR 6285, Lab-STICC Brest, France Deev Interaction, SAS Brest, France Abstract managing simultaneously macro and micro level of control for different entities. Intrinsically, a member of a bio-inspired swarm will have local and environment-driven rules that will be directly impacting on its autonomous navigation function. The level of randomness and the impossibility of controlling the environment itself lead to trajectories that may seem erratic for the human operator, or at least impossible to correlate to highlevel objectives that the operator could be likely to reach and patrol. When controlling “down” at the micro-level of each agent / UAV, the operator is rapidly overwhelmed and cannot provide an adequate amount of attention for managing more than a few UAVs. Moreover, this would mean going back to deterministic plan-driven control, that is out of scope by nature for swarms. The main results issued from our first experiments (Legras et al. 2008; Coppin and Legras 2012) were that the swarm approach seemed to be robust and adapted for simple mission of surveillance, but that the operators in charge of such a system were not ready to understand and dialog with this new kind of system, so that the global performance of the system was potentially spoiled by human intervention. These conclusion is shared by recent work by Kolling et al. who observe that “human operator are generally poor at solving foraging tasks with swarms, not beating the simplest form of autonomy –” (Kolling, Nunnally, and Lewis 2012). Consequently and along with other researchers in the domain (Miner and desJardins 2009), we propose to think about interaction at the swarm level (or at least sub-part of the swarm) rather than at the individual agent’s level within the swarm. Therefore it was proposed to define some intermediary (or transactional) level of interaction that could fill the gap between human and agents representation. This paper presents innovative and intuitive principles of controlling swarms of unmanned vehicles through mission oriented interactions handled on a tabletop. Embedded bio-inspired algorithms are based on the use of artificial pheromones and adapted to the high-level command language mostly based on gestures. Users’ feedback present a synthetic real-time indicator of the current quality of the situation assessment. Preliminary tests results of experiments handled on a scenario of multi-area patrolling are presented. 1 Swarms and UAVs Unmanned Vehicle Systems (UVSs) will considerably evolve within the next two decades. In the current generation of UV Systems, several ground operators operate a single vehicle with limited autonomous capabilities, whereas, in the next generation of UV Systems, a ground operator will have to supervise a system of several cooperating vehicles performing a joint mission, i.e. a Multi-Agent System (MAS) (Johnson 2003; Coppin and Legras 2012). In order to enable mission control, the autonomy of the vehicle and of the system will increase and will require new and richer forms of Human-system interaction. The SMAART project (Legras et al. 2008) demonstrated that interaction between a Human operator and a swarm intelligence is far from being trivial to implement. As this technology shows great promises for empowering large distributed robotic systems, the SUSIE project (SUpervision de Système d’Intelligence en Essaim, supervising swarm intelligence systems) was initiated in order to propose new modes of interaction beyond the simple “mouse clicks – waypoints” paradigm: in a large system the individual robot is probably not the right level of action for an operator, and it may even be superfluous to display each robot’s position. 2 Topology-Based Commands Our approach relies on a topological and geometrical description of objects and actions to be effectuated on these objects. On the algorithm side, these topological features (or constraints) are interpreted in terms of pheromone distribution (masks) or evolution. This has naturally potential indirect effects on the behaviors of the swarm itself. On the operator side, this allows to map operational concepts on traditional geometry defined intuitively through the interaction, and therefore to support a natural and easy way of sending The Need for an Adequate Command Architecture Interacting with swarms gathering a large number of agents (in our case UAVs) strongly emphasizes the dilemma of c 2012, Association for the Advancement of Artificial Copyright Intelligence (www.aaai.org). All rights reserved. 21 orders to the swarms. Table 1 was defined as an elementary taxonomy of topology-based tasks to be supported by a generic swarmbased surveillance or detection system. A set of actions is systematically crossed with a set of topological features in order to explore the space of operational possibilities. The actions that we identified are: The human operator has to express is intent within the available combinations of topological features and actions (e.g patrol a border, avoid a zone), this command is then converted in an input in a form suitable for the corresponding swarm algorithms and send to the swarm for execution. Watch implies for the agents to cover the associated topological feature with their sensors’ field of perception. This coverage should be as thorough (cover the feature in its entirety) and as permanent (cover the feature at every moment) as possible. This action should result in detections i.e. the agents’ sensors registering a certain phenomenon. Project Susie was an effort to test these principles by producing a prototype of a simulated swarm system that would demonstrate very intuitive and innovative interaction principles. The principles of this demonstrator are: 3 Application: Project Susie • use a multitouch tabletop system as a way for the operator to easily express topological commands; Avoid implies to maneuver in order to put as much distance as is reasonable between the agents and the topological feature. This action should be parameterized with a maximum and/or minimum distance. • do not give control of individual agents, but let the operator act at the swarm level (no micro-management). Swarm Characteristics and Missions Find is the action of moving oneself and one’s sensors in order to detect an object in the environment, some of whose characteristics are known (e.g. a red car). This is slightly different from Watch in the sense that an unique or small number of such objects are known to exist and the agents are not waiting for the object to appear, but are actively searching for it. The system selected for Susie consists in a few tens of small, simulated fixed-wing uninhabited aircrafts (15-30 in most of our tests) that are able to perform detection missions under the supervision of a human operator from a tabletop multitouch system. The demonstrator is able to run two kinds of scenarios: • detection and identification of suspicious vehicle convoys on roads or in the open (via EOIR sensors); Follow makes agents align their trajectories with a topological feature. • detection of forest fire (chemical sensors). Intercept makes the agents maneuver so that their trajectories intersect a topological feature as fast as possible. In a typical scenario, the UAVs start in a loitering zone that also acts as a simplified refueling zone. The operator activates or creates operational areas to which some of the UAVs are affected in order to perform detections. These decisions are made by analyzing various sources of intelligence that come from inside and outside the swarm system. When the UAVs are on-site, the system reports to the operator the detections made by the UAVs. IF these detections are confirmed by the operator, he or she his able to position an exclusion corridor that the UAV swarm leave in order to allow another aircraft to intervene safely (e.g. fire-fighting plane). We wanted the operator to be able to deploy the swarm on several zones of interest at the same time, therefore the surveillance of multiple topological features (tasks) can be conducted at the same time. In order to facilitate this — and to further free the operator from the micromanagement of individual vehicles — upon activation or creation of a task, the system computes a nominal subset of the swarm from the loiter zone (if available) and affects these UAVs to said task. The human operator is able to manage the system by requesting transfers of resources (UAVs) between tasks (loiter zone and tasks). UAVs are also able to refuel autonomously by going back to the loiter zone. The topological features that were selected are: points, lines (open or closed), areas and mobile objects (actually a mobile point, one could consider a spreading phenomenon like a cloud). This process allowed us to obtain a set of high-level, behaviors that maps to a broad range of operational concepts like e.g. search & rescue, demining, pollution cleaning, border patrolling, tracking. Swarm Algorithms In order to implement in an actual system some of the behaviors identified in the preceding section, one has to select or develop the corresponding algorithm that will enable the swarm (or sub-swarm) to answer to the human operator’s commands. Figure 1 illustrates the architecture of the command channel (one-way). Swarm Algorithms Swarm Behaviors Topological x actions features In order to fulfill these missions, we selected the combinations [1], [2], [3] and [4] as labelled in Table 1. The Watch action was selected rather than the variations of Find ([5] and [6]) because the objective of the Human operator is to Figure 1: High-level view of the swarm system control architecture. 22 Table 1: Preliminary taxonomy of topology-based tasks for a generic (swarm-based) surveillance system. A brief description of all combinations of actions and topological features is given when applicable. Some combinations sport a numeric label in order to be referred to in the text. Intercept Follow Find Avoid Watch Mobile Point Open line Bear regularly Bear regularly one's Bear regularly one's one's sensors on a sensors on every points of sensors on every points of fixed point in a given line. a closed line. space. 1 2 See Follow Mobile. E.g. collision avoidance, fleeing a threat, or not disturbing another agent. Often associated with a minimum security distance. Do not cross a given line in the environment. One can circumvent the line. See Closed line, with line reduced to a single point. The position of the line can be known in advance or detected by the agents. One has to have access to the position of the mobile via sensors or external intelligence. Find one or several mobile targets in the environment using one's sensors. Intel about the targets' whereabouts or behavior (e.g. direction of travel, speed) can be used. Closed line Reduced case of Mobile. Simpler strategies can be used e.g. systematic search of the environment. Can be associated with a minimum security distance. Maneuver so as to cross a mobile's trajectory in order to carry an action upon it. Implies some knowledge about the mobile's whereabouts (sensors or intelligence). Intel about the geometry of the line can be useful (e.g. length, orientation). The position of the line can be known in advance or detected by the agents. Similar to Open line with the additional dimension of outside/inside. Bear regularly one's sensors on every points of an area. 3 See Closed Line. If the agent happens to be in the area, it has to leave as soon as possible. 4 See Closed line. 6 5 Maneuver so as to keep one's sensors on a mobile target. Find at least one point of a line known to exist in the environment. Do not cross a given closed line (border) in the environment. Equiv. to not leaving/entering an area. Area N/A N/A Maneuver so as to align one's trajectory with a line (oriented or not). Position of the line is known in advance. No sensing activity. Maneuver so as to meet an open line as fast as possible i.e. find closest point. 23 See Open line. Leads to repetitive trajectories e.g. holding pattern. N/A See Open line. N/A deploy its system on different parts of the environment in order to detect possible targets rather than to look for a particular target that is known to exist somewhere. The behaviors that the swarm system is able to exhibit are the following: D • watch lines in the environment i.e. roads that may be used by suspicious vehicle convoys [2]; C • watch pre-defined zones or user-defined zones [3]; B E • watch user-designated points within these two kinds of features i.e. the human operator needs a fly-by of a particular point of interest, for example to confirm a detection; A • avoid a user-positioned air corridor so as to let another aircraft intervene safely [4]. Figure 3: Schematics of the main user interface components: A is the loiter zone where the UAVs are kept in standby and refuel; B is the air corridor that allows other aircrafts to intervene; C is a user-defined active patrol zone; D is a predefined patrol zone that is not yet activated; E is a linear patrol zone. Human Interactions Our whole approach is dedicated to natural and intuitive use of swarms for missions such as patrolling of tracking targets within an area. As explained here below, we propose to found this intuitive operations on topological features and actions that can influence the swarm algorithm and adapt swarm behavior consequently. But these topological feature and objects have also to be determined, defined and pointed at by the operator. Supporting this action by a natural gesture-based interaction seems therefore completely intuitive. We developed a set of simple gestures allowing to draw areas, contours, lines directly on a geographical map displayed on an interactive tabletop system (see Figure 3). While giving a minimal thickness to a line or a contour drawn through a one-finger gesture (see Figure 2), we allow UAVs to consider these objects as special areas and we can manage them with the same pheromone mechanisms. The actual demonstrator was deployed an interactive tabletop DT107 from CircleTwelve with an SXGA+ projector, illustrated on Figure 4. Figure 4: Interactive tabletop setup. Figure 2: Linear zone (bottom) created via a gesture with one finger (top). Algorithms For points of interest, we propose to the operator to tap on the desired location, the more tap the more UAVs to be sent towards the point. In this case, we use the hybrid principle again, so that the point becomes a way-point for the attached UAVs until they reach it, then switching back to pheromonedriven patrol mode. In addition, the user can use simple transfer gestures between zones in order to manage its UAVs assets according to his or her mission-driven constraints and priorities. The agents are driven by several prioritized behaviors: 1. avoid the safety corridor; 2. return to loiter zone if low on fuel; 3. visit interest point if tasked to do so by the system; 4. rejoin patrol zone if affected to one and is outside of it; 5. either: 24 (a) patrol the zone or; (b) stay in the loiter zone. Behavior 1 is achieved by calculating a simple escape heading when the UAV is too close to the corridor. Behaviors 2, 3 and 4 are achieved by computing an appropriate waypoint and assigning to the UAV. Behavior 5b is a custom circling behavior developed for demonstration purposes. Behavior 5a is achieved via a pheromone grid-based algorithm. Figure 5 illustrates how real-world coordinates are mapped to the pheromone grid that supports the algorithm. Given a cell-size, the smallest enclosing grid for the region of interest is computed, as well as a mask to constrain the behavior to the arbitrary shape input by the operator. Maximum levels of pheromone (255, white on the figure) are deposited within the region to patrol. The algorithm follows the two principles of consumption and deposition: • the agents consume (set to zero) pheromone along their movement with respect to their perception radius; • the agents locally determine their trajectory so as to consume the highest levels of pheromone; • fine layers of pheromone (value increments) are regularly deposited so that agents come back to places visited earlier as the pheromone that was consumed is replaced. One can note a difference with other bio-inspired pheromone algorithms that employ deposition, diffusion and evaporation principles. 4 Discussion and Perspectives We have only discussed in this paper the broad functionalities that allow the operator intuitively to issue commands to the swarm. These commands include: designation, creation and suppression of arbitrary-shaped patrol zones (linear and polygonal), designation of interest point for the swarm, positioning of an exclusion corridor and transfer of resources (UAVs) between patrol zones. We have developed a synthetic indicator that allows the system to express the performance status of patrol tasks. This indicator is not detailed in this article, but can be seen on Figures 2 and 5, it consists in a panel that indicates the optimal and current numbers of UAVs for a zone, as well as a composite representation of instantaneous patrol performance, optimality and upper bound. Future work on this topic include working with heterogeneous robots (different UAVs, or mixing ground and air vehicles), developing new tasks that require more complex interactions between the swarm and its tactical environment (e.g. rendezvous) and moving beyond simulation by setting up a live demonstrator. Figure 5: An arbitrary-shaped zone (top) has been created and two UAVs are patrolling. One can see a representation of the rectangular pheromone grid that supports the corresponding swarm algorithm (bottom). References Coppin, G., and Legras, F. 2012. Autonomy spectrum and performance perception issues in swarm supervisory control. Proceedings of the IEEE 100(3):590–603. Johnson, C. 2003. Inverting the control ratio: Human control of large, autonomous teams. In Proceedings of AAMAS’03 Workshop on Humans and Multi-Agent Systems. Kolling, A.; Nunnally, S.; and Lewis, M. 2012. Towards human control of robot swarms. In Yanco, H. A.; Steinfeld, A.; Evers, V.; and Jenkins, O. C., eds., HRI, 89–96. ACM. Legras, F.; Glad, A.; Simonin, O.; and Charpillet, F. 2008. Authority sharing in a swarm of UAVs: Simulation and experiments with operators. In Proceedings of the International Conference on Simulation, Modeling and Programming for Autonomous Robots (SIMPAR 2008). Miner, D., and desJardins, M. 2009. Predicting and controlling system-level parameters of multi-agent systems. In AAAI Fall Symposium on Complex Adaptive Systems and the Threshold Effect. Acknowledgements SUSIE is an Exploratory Research and Innovation contract (Recherche Exploratoire et Innovation: REI, 2009.34.0003) of the French Defense Procurement Agency (Délégation Générale pour l’Armement: DGA, Mission pour la Recherche et l’Innovation Scientifique: MRIS) conducted by Télécom Bretagne and INRIA Loria. 25