DAMN: A Distributed Architecture for Mobile NavigationJulio K. RosenblattRobotics InstituteCarnegie Mellon UniversityPittsburgh, PA [email protected] architecture is presented in which distributed task-achieving modules, or behaviors, cooperatively determine amobile robot’s path by voting for each of various possibleactions. An arbiter then performs command fusion and selectsthat action which best satisfies the prioritized goals of thesystem, as expressed by these votes, without the need toaverage commands. Command fusion allows multiple goalsand constraints to be considered simultaneously. Examples ofimplemented systems are given, and future research directionsin command fusion are discussed.Keywords: mobile robots, distributed architecture, behaviors,voting, arbitration, command fusionIntroductionIn order to function in unstructured, unknown, or dynamicenvironments, a mobile robot must be able to perceive itssurroundings and generate actions that are appropriate forthat environment and for the goals of the robotic system.Mobile robots need to combine information from severaldifferent sources. For example, the CMU Navlab vehiclesare equipped with sensors such as video cameras, laser rangefinders, sonars, and inertial navigation systems, which arevariously used by subsystems that follow roads, track paths,avoid obstacles and rough terrain, seek goals, and performteleoperation. Because the raw sensor data and the internalrepresentations used by these subsystems tend to varywidely, especially when these modules have been developedindependently, combining them into one coherent systemwhich combines all their capabilities has proven to be verydifficult.To function effectively, an architectural framework forthese sensing and reasoning processes must be imposed toprovide a structure with which the system may be developed,tested, debugged, and understood. However, the architectureshould serve as an aid, not a burden, in the integration ofmodules that have been developed independently, so it mustnot be overly restrictive. It must allow for purposeful goal-oriented behavior yet retain the ability to respond topotentially dangerous situations in real-time whilemaintaining enough speed to be useful. The earliest work in robot control architectures attemptedto reason by manipulating abstract symbols using only purelogic (Nilsson, 1984). The limitations of this top-down AIapproach led to a new generation of architectures designed ina bottom-up fashion to provide greater reactivity to therobot’s surroundings, but sacrificed generality and the abilityto reason about the system’s own intentions and goals. Hierarchical approaches allow slower abstract reasoning atthe higher levels and faster numerical computations at thelower levels, thus allowing varying trade-offs betweenresponsiveness and optimality as appropriate at each level(Payton, 1986; Albus, McCain and Lumia, 1987). Whilesuch an approach provides aspects of both deliberativeplanning and reactive control, the top-down nature ofhierarchical structures tends to overly restrict the lowerlevels (Payton, Rosenblatt and Keirsey, 1990). Inhierarchical architectures, each layer controls the layerbeneath it and assumes that its commands will be executedas expected. Since expectations are not always met, there is aneed to monitor the progress of desired actions and to reportfailures as they occur (Simmons, Lin and Fedor, 1990). In anunstructured, unknown, or dynamic environment, thisapproach introduces complexities and inefficiencies whichcould be avoided if higher level modules participated in thedecision-making process without assuming that theircommands will be strictly followed.Rather than imposing a top-down structure to achieve thisdesired symbiosis of deliberative and reactive elements, theDistributed Architecture for Mobile Navigation takes anapproach where multiple modules concurrently share controlof the robot by sending votes to be combined rather thancommands to be selected and executed (Rosenblatt, 1995;Rosenblatt and Thorpe, 1995).The Distributed Architecture for Mobile Navigation hasbeen successfully used to integrate the various subsystemsmentioned above, thus providing systems that perform taskssuch as road following, cross-country navigation, andteleoperation while avoiding obstacles and meeting missionobjectives. In addition to its use on the CMU Navlabvehicles, DAMN has also been used at Martin Marietta,Hughes Research Labs, and the Georgia Institute ofTechnology. The Distributed Architecture forMobile NavigationDeliberative planning and reactive control are equallyimportant for mobile robot navigation; when usedappropriately, each complements the other and compensatesfor the other’s deficiencies. Reactive components providethe basic capabilities which enable the robot to achieve low-level tasks without injury to itself or its environment, whiledeliberative components provide the ability to achievehigher-level goals and to avoid mistakes which could lead toinefficiencies or even mission failure. But rather thanimposing an hierarchical structure to achieve this symbiosis,the Distributed Architecture for Mobile Navigation (DAMN)takes an approach where multiple modules concurrentlyshare control of the robot. In order to achieve this, a schemeis used where each module votes for or against variousalternatives in the command space based on geometricreasoning, without regard for the level of planning involved.Figure 1 shows the organization of the DAMNarchitecture, in which individual behaviors such as roadfollowing or obstacle avoidance send votes to the commandarbitration module; these inputs are combined and theresulting command is sent to the vehicle controller. Eachaction-producing module, or behavior, is responsible for aparticular aspect of vehicle control or for achieving someparticular task; it operates asynchronously and in parallelwith other behaviors, sending its outputs to the arbiter atwhatever rate is appropriate for that particular function. Eachbehavior is assigned a weight reflecting its relative priorityin controlling the vehicle. A mode manager may also be usedto vary these weights during the course of a mission based onknowledge of which behaviors would be most relevant andreliable in a given situation.DAMN is a behavior-based architecture similar in someregards to reactive systems such as the SubsumptionArchitecture (Brooks, 1986). In contrast to more traditionalcentralized AI planners that