From virtual world to reality:designing an autonomous underwater robot
Donald P. BrutzmanComputer Science Department, Naval Postgraduate School
Monterey California 93943-5000 USAbrutzman@cs.nps.navy.mil
(408) 646-2149 work, (408) 646-2595 fax
AbstractDesign of autonomous underwater robots is particularlydifficult due to the physical and sensor challenges ofthe underwater environment. Inaccessibility duringoperation and low probability of failure recovery makesrobot stability and reliability paramount. Building anaccurate and complete virtual world simulation isproposed as a necessary prerequisite for design of anautonomous underwater robot. A virtual world caninclude actual robot components and models for allother aspects of the world. Robot design can be fullytested using a virtual world and then verified using thereal world. Additional testing can be performed in thevirtual world that is not feasible in the real world.Visualization of robot interactions within a virtual worldpermits sophisticated analysis of robot performance thatis otherwise unavailable. All aspects of world modelingand robot design must be mastered and coordinated inorder to build an authentic virtual world and capableautonomous robot.
1 ProblemsAutonomous underwater robot design is difficult.Unlike most other mobile robots, underwater robotsmust operate unattended and uncontrolled in a remoteand unforgiving environment. Inaccessibility duringoperation greatly complicates the design and evaluationof system software. In order to ensure completereliability, however, robot software and hardware needto be fully tested in a controlled environment beforeoperational deployment. Such comprehensive testingrequirements cannot be met using a standalonelaboratory robot due to the complexity andunpredictability of interactions that can occur in theactual remote environment. A different approach isneeded which can effectively support research on themany problems facing underwater robot designers.
Underwater robots are normally termedautonomous underwater vehicles (AUVs), not becausethey are intended to carry people but rather becausethey are designed to intelligently convey sensors andpayloads. AUVs must accomplish complex tasks and
Submitted to the AAAI Fall Symposium on Applications of ArtificialIntelligence to Real-World Autonomous Mobile Robots, Cambridge,Massachusetts, October 23-25, 1992.
diverse missions while maintaining stable physicalcontrol with six spatial degrees of freedom. Little or nocommunication with distant human supervisors ispossible. When compared to indoor, ground, airborneor space environments, the underwater domain typicallyimposes the most restrictive physical control and sensorlimitations upon a robot. Underwater robot designrequirements therefore motivate this examination.Considerations and conclusions remain pertinent asworst-case examples in other environments.
A large gap exists between the projections oftheory and the actual practice of underwater robotdesign. Despite a large number of remotely operatedsubmersibles and a rich field of autonomous robotresearch results (Iyengar and Elfes 1990a, 1990b), fewAUVs exist and their capabilities are limited. Cost,inaccessibility and scope of AUV design restrict thenumber and reach of players involved. Interactions andinterdependencies between hardware and softwarecomponent problems are poorly understood. Testing isdifficult, tedious, infrequent and potentially hazardous.Meaningful evaluation of results is hampered by overallproblem complexity, sensor inadequacies and humaninability to directly observe the robot in situ. Potentialloss of an autonomous underwater robot is generallyintolerable due to tremendous investment in time andresources, likelihood that any failure will becomecatastrophic and difficulty of recovery.
Underwater robot progress has been slow andpainstaking for many reasons. By necessity mostresearch is performed piecemeal and incrementally. Forexample, a narrow problem might be identified assuitable for solution by a particular artificialintelligence (AI) paradigm and examined in great detail.Conjectures and theories are used to create animplementation which is tested by building a model orsimulation specifically suited to the problem inquestion. Test success or failure is used to interpretvalidity of conclusions. Unfortunately, integration ofthe design process or even final results into a workingrobot is often difficult or impossible. Lack ofintegrated testing prevents complete verification ofconclusions.
AUV design must provide autonomy, stabilityand reliability with little tolerance for error. Controlsystems require particular attention since closed-form
From: AAAI Technical Report FS-92-02. Copyright 1992, AAAI (www.aaai.org). All rights reserved.
solutions for many hydrodynamics control issues areunknown. In addition, AI methodologies are essentialfor many critical robot software components, but theinteraction complexity and emergent behavior ofmultiple interacting AI processes is poorly understood,rarely tested and impossible to formally specify. Betterapproaches are needed to support coordinated research,design and implementation of underwater robots.
Despite these many handicaps, the numerouschallenges of operating in the underwater environmentforce designers to build robots that are truly robust,autonomous, mobile and stable. This fits well with amotivating philosophy of Hans Moravec: ".. solvingthe day to day problems of developing a mobileorganism steers one in the direction of generalintelligence... Mobile robotics may or may not be thefastest way to arrive at general human competence inmachines, but I believe it is one of the surestroads." (Moravec 1983)
2 Multiple Interacting ProcessesDesigning an AUV is complex. Many capabilities arerequired for an underwater mobile robot to act capablyand independently. Stable physical control, motioncontrol, sensing, motion planning, mission planning,replanning and failure recovery are example softwarecomponents that must be solved individually fortractability. The diversity and dissimilarity of thesemany component subproblems precludes use of a singlemonolithic AI paradigm.
Distributed AI usually addresses specificationsand protocols between similar autonomous agentsworking cooperatively on global problems. Hybridreasoning often refers to novel combinations of two orthree techniques to improve overall performance whensolving a single problem type. Neither definitionappears suitable for general robot control. Multipledissimilar AI processes must interact in an intelligentmanner to achieve the robust capabilities and multiplebehaviors needed by a mobile robot (Elfes 1986). Avariety of robot architectures have been proposed anddeveloped to provide the control framework underwhich multiple AI processes can interact. A briefdiscussion of current robot architectures is thereforeuseful to clarify the scope of robot design issues.
Robot architectures can be classified over aspectrum that ranges from hierarchical to reactive(Byrnes et al. 1992). Hierarchical architectures aredeliberative, symbolic, structured, "top down,"goal-driven, have explicit focus of attention and areoften implemented using backward inferencing.Hierarchical approaches typically contain world modelsand use planning and search techniques to achievestrictly defined goals. Hierarchical architectures tend tobe somewhat rigid, unresponsive in unpredictedsituations and computation-intensive, yet remain capableof highly sophisticated performance.
Reactive architectures are subsumptive,"bottom up," sensor-driven, layered and may often becharacterized by forward inferencing. Reactivearchitectures attempt to combine robust subsumingbehaviors while avoiding dynamic planning and worldmodels. Reactive architectures appear to behavesomewhat randomly and achieve success withoutmassive computations by using well-consideredbehaviors that tend to lead to task completion(Brooks 1986). Scaling up to complex missions isdifficult. Stability and deterministic performance iselusive.
It is interesting to note that numerous robotarchitecture researchers have recently proposed hybridcontrol architectures (Kwak et al. 1992) (Bonasso et al.1992) (Bellingham and Consi 1990) (Payton and Bihari1991) (Spector and Hendler 1991). A common themein these proposals is integrating the long-termdeliberation, planning and state information found inhierarchical approaches with the quick reaction andadaptability of subsumptive behaviors. Individualweaknesses of hierarchical and reactive architecturesappear to be well-balanced by their respective strengths.
Stability and reliability deserve repeated mentionin the context of multiple interacting processes. Controlsystem considerations are often overlooked under theguise of simplifying assumptions that hide importantreal world restrictions and pitfalls. Robot survivabilitydictates that physical and logical behavior must alwaysconverge to a stable yet adaptive set of states.Divergence, deadlock, infinite loops and non-lineardynamic behavior must be detectable and controllable.Real-time operating constraints on sensing, processing,action and reaction must be similarly resolved.Stability prerequisites become similarly important forground robots as they progress from structured tounrestricted environments.
3 Virtual WorldThe broad requirements of underwater robot designprovide a strong argument against piecemeal designverification. Individual component simulations are notadequate to develop effective AI-based systems orevaluate overall robot performance.
Virtual world systems provide the capability tosee and interact with distant, expensive, hazardous ornon-existent three-dimensional environments (Zyda andPratt 1992). A virtual world is intended to providecomplete functionality of the target environment in thelaboratory. A virtual world can provide adequatesimulation scope and interaction capability to overcomethe inherent design handicaps imposed when building aremote robot to operate in a hazardous environment.Construction of a virtual world for robot developmentand evaluation is hereby proposed as a necessaryprerequisite for successful design of a complex remoterobot such as an AUV.
From: AAAI Technical Report FS-92-02. Copyright 1992, AAAI (www.aaai.org). All rights reserved.
A virtual world which is used to recreate everyaspect of the environment external to the robot mustalso include robot sensors and analog devices (such asthrusters and rudders) which are impossible torealistically operate in a laboratory. Interactionsbetween software processes, vehicle hardware and thereal world must all be comprehensively modeled andmutually consistent. Robot physical behavior andsensor interactions must be adequately simulated. Therobot itself is directly plugged into the virtual worldusing normal sensor and actuator connections. Thedifference between operation in a virtual world or anactual environment must be transparent to the robot inorder to be effective. Successful implementation of thevirtual world can be validated by identical robotperformance in each domain.
The potential value of general world simulatorshas been previously recognized (Moravec 1988).Certain underwater robots already benefit from theavailability of capable simulators (Pappas et al. 1991)(Brutzman 1992a) (Brutzman et al. 1992b). However,once simulation sophistication reaches that of a virtualworld, interesting things become possible. Emergentbehavior from interaction between multiple AIprocesses and the environment becomes evident.Sensor interactions can be repeated indefinitely in orderto develop new analysis algorithms and achievefine-tuned sensor performance. Machine learning basedon massive repetitive training becomes feasible and canbe conveniently monitored. Potentially fatal scenarioscan be attempted without risk to robot, human orenvironment.
A plethora of interrelated requirements requiresmastering all aspects of world modeling and robotdesign in order to build both an authentic virtual worldand a capable autonomous robot. Traditionally onlycomponent software processes were integrated into therobot architecture; now corresponding validationsimulations must also be integrated into the virtualworld. Adequate simulation of a comprehensiveunderwater virtual world is possible since precisemodels are available for kinematics (Badler et al. 1991)(Zyda et al. 1991) (Zyda et al. 1990), hydrodynamics(Yuh 1990) (Abkowitz 1969), sonar response (Etter1991), and other objects in the underwater environment(Pentland 1990). Characteristics of the underwatervehicle physical components which require modelingcan be handled on a case basis (Pappas 1991).Depending upon planned missions and AUV hull form,interactive terrain modeling (Pratt et al. 1992) andprecise hydrodynamics modeling are areas likely torequire further investigation prior to implementation.Virtual world construction requires significantcoordination and cross-disciplinary efforts.
From a design viewpoint, models replacefunctionality otherwise found only in the real world.Identical replication of every world characteristic is
impossible but partial modeling of all pertinent worldaspects is necessary. Exact reproduction of real worldbehavior might even be undesirable in some cases. Forexample, a sonar model might provide perfect rangeand intensity values that can be augmented bystatistically adjustable noise and errors. Initial trainingor evaluation of a new sensor algorithm is bestperformed using the "perfect" error-free model. Furtheranalysis using quantifiable noise and errors will providenew insight into algorithm robustness and adaptability.Because it has been coupled with the robots ability tomove freely throughout a virtual world, overalleffectiveness of this example AI-based sensor processis likely to be superior to that obtainable by training inthe real world.
Models in virtual worlds can be further studiedto examine the benefits of controllable error precision,addition of uncertainty, incorporation and extension ofactual sensor data, and maintaining internal worldconsistency throughout mutual interactions betweenindividual component models and the multifacetedautonomous robot. Such study also clarifiesunderstanding of robot design problems andspecifications.
4 VisualizationA robot interacting within a virtual world permitscomplete visualization of all aspects of both robot andvirtual environment. Visualization of robotperformance is essential for evaluating both the precisedetails of low-level execution and the broad suitabilityof high-level behaviors. Visualization of robotinteractions permits sophisticated analysis that is notpossible using traditional test methods such asindividual software module evaluation, direct robotobservation or post-mission reconstruction.
Human beings are visually oriented. Being ableto see and control location and time in a moving pictureallows us to quickly and intuitively understand...