AI Magazine Volume 27 Number 1 (2006) (© AAAI) Articles Components, Curriculum, and Community: and in Undergraduate AI Education

Zachary Dodds, Lloyd Greenwald, Ayanna Howard, Sheila Tejada, and Jerry Weinberg

■ This editorial introduction presents an of the engaging and active communities that overview of the robotic resources available to support robotic competitions and exhibitions. AI educators and provides context for the arti- These -based components, curricula, and cles in this special issue. We set the stage by ad- communities, we hope, broaden the resources dressing the trade-offs among a number of es- available to educators, as we all invite students tablished and emerging hardware and software to share our enthusiasm for AI. platforms, curricular topics, and robot contests used to motivate and teach undergraduate AI. Robot Platforms for AI Education Stuart Russell and Peter Norvig frame their widely used AI text through a paradigm of in- obot platforms have played a fundamen- telligent agents (Russell and Norvig 2003). tal role in the field of artificial intelli- Such an approach resonates with students, all Rgence (AI) for more than 30 years. Yet it of whom have deep experience with (and as) is only recently that physically embodied intelligent agents. Yet nearly all of that experi- agents have become a viable tool in the under- ence is with embodied intelligent agents, and graduate AI classroom. Examples of the flurry this familiarity makes robots a strong motiva- of activity in this area include competitions tor of AI. This embodiment contrasts with the and exhibitions, the growing options for low- majority of computer science subfields, in cost robot hardware and software, and a num- which computers interact with the physical ber of recent workshops and symposia. This world very differently than we do. What’s special issue of AI Magazine grew out of the more, for AI educators, robotic hardware is not 2004 AAAI spring symposium on Accessible, only a hook that can draw students to the field, Hands-on AI and Robotics Education. In this but a fundamental facet of the AI endeavor. article, we seek to showcase how robots have The challenge is to find a set of hardware and influenced both the curriculum and practice of software resources that serve both as motiva- teaching AI at the undergraduate level. tion and as tools to advance, not limit, the AI This survey article first overviews a number that students pursue in an academic course of of robot platforms and presents trade-offs in study. choosing among them. We then highlight the Today there exists a large and growing selec- variety of AI curricula supported by low-cost ro- tion of robotic platforms suitable for conveying botic platforms. We conclude with a summary and investigating fundamental AI topics. Fig-

Copyright © 2006, American Association for Artificial Intelligence. All rights reserved. 0738-4602-2006 / $2.00 SPRING 2006 11 Articles

ures 1, 2, and 3 summarize some of these re- Nintendo Game Boy Advance’s Xport Robot sources and their capabilities, with particular Controller (XRC) and related XBC (LeGrand et attention given to newer models and those al. 2005). Any device with a serial port or a con- widely employed at the undergraduate level. version to one can drive up to eight servo mo- These sensors are available from a number of tors using Pontech's SV 203, which is a small, retailers including HiTechnic Products and easily programmable controller (Bishop et al. Mindsensors.com for the RCX-compatible sen- 2004). Though less well established than the sors and Acroname for those sensors not specif- Handy Board and the Lego RCX controllers, all ically tailored to the RCX’s Lego interface. of these systems have been employed to teach Before discussing the platforms listed in fig- undergraduate AI and/or robotics. They typi- ure 1, it is worth mentioning a family of low- cally use a Lego or custom chassis and require cost robotics resources we have omitted: those some work to interface with Lego sensors. Oth- dedicated to teaching the electrical and me- er prebuilt controllers are also available, such as chanical engineering that underlies most con- Ridgesoft’s IntelliBrain, a more powerful, pro- temporary robotics, such as the basic stamp mi- prietary alternative to the Handy Board. crocontroller (Kuhnel and Zahnert 1997). One perceived disadvantage of robot kits is Undergraduate AI does not ignore the impact that the resulting platforms can provide soft- of such design decisions but instead focuses on ware support for only a low level of behavioral the computational challenges those decisions abstraction. Recent example curricula, such as create. In the context of AI education the hard- Greenwald’s (see his article in this issue of AI ware/software interface, that is, the ease with Magazine), mitigate these concerns: topics as which students can interact computationally computationally demanding and subtle as A* with a robot and investigate how their algo- search and Markov decision processes have run rithms behave, is a crucial criterion for evaluat- entirely on a Handy Board. Further, because it is ing robotic platforms. easy to download information from each of these platforms to a PC, all of these systems can Hardware and Software be used to collect data for off-board analysis, for A key advantage of the two most popular plat- example, learning the weights of a back propa- forms, Lego Mindstorms (or RCX brick) and the gation or Bayesian network (Greenwald and Handy Board, is the variety of ways in which Artz 2004). Susan Imberman demonstrated that students can program them. C-like languages the results of such analysis can return to the ro- and Java subsets are available for the Mind- bot for control, such as in a line-following task storms through the BricxCC and LeJOS based on neural network parameters learned off firmware upgrades. Both are open-source pro- board (Imberman 2004). On-board processing is jects with substantial deployment. Interactive needed only to the extent that sensed data must C is the default computational interface on the contribute to behavioral decisions during oper- Handy Board. A commercial Java implementa- ation. In practice, a more important disadvan- tion, RoboJDE, is available for the Handy Board tage of all of these platforms’ flexibility is that from RidgeSoft, LLC.1 These two platforms’ they are more difficult to support with simula- large user communities breed support for a tion software, as neither the robot morphology wide variety of interfaces: of particular note is nor its sensor suite is known a priori. the Lisp interface to the Lego RCX brick de- In contrast to robot kits, preassembled plat- scribed in detail later in this AI Magazine issue. forms offer additional capabilities and conve- Both Lego and Handy Board platforms provide nience at a higher cost. K-Team’s miniature ro- a microcontroller to which students attach a bot, the Khepera, provides options for a huge chassis, motors, and sensors. Their flexibility variety of sensors and actuators including color enables students’ hands-on investigation of the vision and a parallel-jawed gripper. Its small close relationship between physical agents’ form factor facilitates use in almost any space: form and function. Fred Martin’s text Robotic students have watched a Khepera execute their Explorations: An Introduction to Engineering programs from a desktop in a professor’s office Through Design (Martin 2000) is a popular and (Challinger 2005). The Khepera has proven it- natural starting point for Handy Board–based self within AI classes at a variety of schools courses. Several texts also build curricula (Harlan 2005, Kumar and Meeden 1998). K- around the RCX, such as Bagnall’s Core Lego Team’s newer and slightly larger robot, the He- Mindstorms Programming (Bagnall 2002). mission, also uses a ring of infrared (IR) sensors Other modular robotics options have grown for proximity sensing—at a cost an order of around off-the-shelf computational engines: magnitude lower than the Khepera. Not as ex- the Palm Pilot (Reshko, Mason, and pandable, it has also not yet received the same Nourbakhsh 2002; Avanzato 2004) and the scrutiny in educational or research settings.

12 AI MAGAZINE Articles

Lego Mindstorms Popular because it works out-of-the-box, the yellow lego RCX brick has inspired curricula and freely $199 10x3x6 cm (w/o chassis) available software, such as is described in Klassner (2004), and Parsons and Sklar (2004). Sonar, rota- www.legomindstorms.com tion, and IR sensors can extend provided touch/ light inputs.

Xport Robot Controller The Xport leverages the computation of a hand- held GameBoy (not included). KIPR offers a very $269 8x3x1cm (w/o chassis) capable vision-augmented system, the XBC (Le- Grand et al. 2005), that provides easily confi- www.charmedlabs.com gurable and very powerful region tracking.

Handy Board Controller Although the price does not include sensors or a typically Lego chassis, the Handy Board supports touch, light, sonar, IR, compass, and $299 12x8x3 cm (w/o chassis) vision sensing. It is used in many AI courses (Imberman 2004, Danyluk 2004, Martin and handyboard.com Pantazopoulos 2004).

The Palm Pilot stand-alone kit relies on Palm Pilot Robot Kit Acronameís B rainStem controller, three IR ranging sensors, and a handheld computer. $315 18 cm dia. x 6 cm Designed at Carnegie Mellon University (Reshko, Mason, and Nourbakhsh 2002) the PPRK has been www-2.cs.cmu.edu/~pprk employed for outreach beyond AI robotics (Avanzato 2004).

Robix/Pontech The Robix enables many forms Manipulator via six servos and connective hardware. The Pon- tech SV203 ($60), an 8-motor serial controller, $550 varying dimensions can also control these servos (Crabbe 2004, Bishop et al. 2004, Sutherland www.robix.com 2000).

Used in the RoboCup legged league, the AIBO Sony AIBO offers a microphone, vision, and touch sensing with 802.11b wireless for communications. $1900 10x3x6 cm (w/o chassis) A number of freely available software interfaces exist, as well as tested AI robotics curricula www.sony.net/Products/aibo (Veloso et al. 2004).

ActivMedia Robots The Amigobot at right is the least expensive of a large line of prebuilt robots from ActivMedia. $2000+ 10x3x6 cm With an RF link, sonar ring, odometry, and an interface to the ActivMedia Aria simulator, this robot spans research and educational uses www.activmedia.com (Konolige et al. 2004, Arkin 2000).

K-Team Robots The Khepera ($2025+), the 5-cm platform at left, has been used in several AI robotics courses (see, $296-2025+ 5-12cm dia. x 5cm for example, Harlan [2005]). It accepts a huge array of sensors at commensurate www.k-team.com costs. The new Hemisson robot (at right) is $296 but untested.

Figure 1 A Comparison of the Price, Form Factor, Sensing, and Computational Capabilities of Eight Low-Cost Robotic Platforms Used in AI Robotics Settings. Product-specific URLs are on the left; references to example uses in undergraduate settings are included in the notes on each platform.

SPRING 2006 13 Articles

Sony’s robotic quadruped, the AIBO, is support- is not sonar. Coping with this uncertainty has ed by several freely available software develop- long been a motivation driving the field of AI ment environments. Veloso and Rybski’s article Robotics. Not every simulator models sensing in this AI Magazine issue attests to the breadth and actuation as noisy processes: Stage, de- and depth of AI topics that the AIBO can sup- signed for low-fidelity simulation of many port. At approximately US$2000, ActivMedia’s agents, does not; Gazebo does. Amigobot is about as expensive as an AIBO. It Perhaps the ideal educational toolset com- is a more traditional wheeled platform with a prises both a simulator and robot hardware ring of sonar rangers providing the primary with identical programming APIs. Player, a ro- sensory input. One advantage of such a larger, bot server that accompanies Stage and Gazebo, stronger platform is that existing resources may offers drivers that interface with many robots be easier to integrate. For example, any robot and sensors, though not typically the robotic- that can support a laptop or handheld comput- kit platforms. Webots allows compilation of er can support on-board vision, global posi- control programs to Lego’s RCX brick, the AI- tioning (GPS), and a host of other off-the-shelf BO, and K-Team’s robots. Pyro offers a single inputs. streamlined interface to many simulated and physical robotic platforms. Other simulators Simulation offer API support for their own hardware. If robotic platforms’ primary contribution to AI Such tandem systems allow educators to pro- is their computational interaction with the vide access to robotic experiments without world, why not abstract away the hardware stocking a large lab full of platforms. Students completely? Simulation is an attractive option develop and test software in simulation before for AI educators for many reasons: the speed of transferring their code for trial runs on one or students’ design-test-debug cycles, the repeata- a small number of physical robots. Though bility of experiments, the availability of rich vi- breakdowns and scheduling conflicts are po- sualization tools that may not be available on tential concerns with fewer robots, ease of board a physical platform, and the (simulated) maintenance and a natural reinforcement of access to expensive or unavailable platforms or the software design and testing process are cer- sensors. Gazebo, for instance, can simulate tainly plusses. Hardware slows down students’ both laser range finders and autonomous heli- code-test-debug loops dramatically. Depending copters. on the extent to which realistic sensor noise Many robot simulators support a particular and other sources of uncertainty are modeled platform or controller, typically higher-end sys- in simulation, students’ programs may require tems. The Aria simulator supports ActivMedia’s substantial rewriting in order to succeed on line of robots; Webots grew from the Khepera physical hardware. Such experiences, though robot’s simulator and works equally well with frustrating, also offer the opportunity for deep- the newer Hemisson. Others have evolved for er insight into the difficulty of computational and from specific tasks, for example, RoboCup interaction with the physical environment. soccer and rescue league simulators. Spending a fixed budget on fewer platforms Two powerful, general-purpose simulators may also better serve undergraduate research are the open-source projects Stage and Gazebo projects by providing a richer sensor suite than (Vaughan, Gerkey, and Howard 2003), which a large number of lower-cost kits can. Simula- support research-level work in AI and robotics. tors can then expand the reach of a few (or no) Having grown out of the Interaction Lab at the hardware resources to a much larger class of University of Southern California, these are ac- students. tively supported tools that provide access to a rich set of simulated sensing modalities: vision, Sensing and Computation laser scans, sonar, and global positioning. Fig- The sensor suite available is likely to have the ure 2 contrasts the purposes and capabilities of greatest curricular—and financial—impact on four robot simulators. As detailed later in this an undergraduate’s experience with AI robot- issue, the Pyro project creates an interface that ics. To a large degree, it is the richness and eases access to these powerful tools within ed- reach of a platform’s sensors that drive both its ucational settings. cost and its capabilities. Student projects with One drawback to using simulation is the loss inexpensive robot kits tend to focus on local of the physical embodiment that attracts many sensing: contact sensors that detect collisions, students to learning AI with robots. Another photoresistors for determining ambient light potential problem is the loss of the unpre- strength, and short-range infrared distance sen- dictability of real-world physical interaction. sors. Reactive architectures are natural for these Simulated vision is not vision; simulated sonar platforms, localization and navigation much

14 AI MAGAZINE Articles

Stage/Gazebo Stage offers multiagent 2D simulation; Gazebo provides more fidelity in 3D simulation Open-source software of a few robots (Vaughan, Gerkey, and Howard 2003). This image shows simulated vision on a Pioneer. The Player system connects Stage/Gazebo playerstage.sourceforge.net to many physical robots.

Robocup Soccer The RoboCup simulation league runs competitions Simulator atop the original 2D simulator (this image is a Pyro client), a newer 3D version, and for Open-source software contributions to its codebase. It is the backdrop of sserver.sourceforge.net several AI courses (Stone 2004, Coradeschi and Malec 2000).

Robocup Rescue The RoboCup rescue competition began in 2001 Simulator and includes physical-robot and simulated leagues. This simulator models intact/collapsed Open-source software structures, cars/traffic flow, fire progress, civilians, sourceforge.net/projects and emergency crews in 2D and 3D (Kleiner 2004) /roborescue

Cyberbotics Webots Like Stage/Gazebo, this physics-based simulator offers drivers for physical platforms common to $2750 for 10 licenses undergraduate education: Aibo, RCX, and K-Team robots. In contrast to those systems, Webots is also www.cyberbotics.com supported on Windows systems (Michel 2004).

Figure 2. An Overview of the Capabilities and Costs of Four Robot Simulators. less so. However, as illustrated in figure 3, access to spatial reasoning algorithms from ba- sonar, IR rangers, and color vision are available sic wall following to topological mapping. En- as extras for the Lego RCX and Handy Board coders are available, too, for the Handy Board, platforms (though the Lego Vision Command RCX, XBC, and the Intellibrain. The XBC fur- system camera is tethered to a personal com- ther computes position and velocity by mea- puter). Relatively recently, the Kiss Institute for suring the motors’ back-emf, obviating the Practical Robotics2 has offered the XBC, a pow- need for encoders. All of these smaller con- erful controller for Lego platforms (LeGrand et trollers also support a sonar or IR sensor al. 2005). The XBC can support an extensive mounted on a rotating servomotor turret as a sensor suite, and its support for vision is partic- less expensive alternative to a ring of range sen- ularly strong: an integrated camera, the capa- sors. Vision facilitates landmark detection and bility for multiple-region tracking, and the very identification, along with the classification and accessible configuration of its image process- deliberative tasks that can build on those capa- ing. Like Charmed Labs’ XPort Robot Con- bilities. Although even simple platforms can troller on which the XBC is based, these emerg- motivate research-level AI robotics projects ing resources leverage the powerful processor (Huang and Beevers 2004), it is primarily the and color screen available in the ubiquitous sensing available that provides options to stu- handheld Game Boy Advance. dents, educators, and researchers alike. Shaft encoders, which allow the measure- Sony’s AIBO robotic dog stands out in a ment of position and velocity, and a ring of number of ways among the platforms in figure range sensors (sonar in the case of the Pioneers, 1—it is the only and it offers an ar- IR for the Khepera and Hemisson) offer student ray of sensors and computational facilities rich

SPRING 2006 15 Articles

enough to support deliberative, cooperative tation ubiquitous. As of October 2005, the tasks like soccer. The AIBO user community, Roomba offers a serial interface to interested ro- represented in this issue by Manuela Veloso boticists.4 The extent to which communities and Paul Rybski’s article, has created a set of will develop to support these emerging plat- software resources that make the robot a forms with software, sensors, and curricula still promising one for AI education. Their abstrac- remains an open question. tions of low-level behaviors and raw sensor in- In addition, the impact, marketing power, put make the AIBO particularly suitable for in- and economies of scale in the toy and enter- vestigating task-directed decision making in tainment industry will continue to play impor- the face of uncertainty. tant roles in creating inexpensive robotics re- The computational resources among the sources. The use of the AIBO for research and platforms in figure 1 vary widely, and they can education has followed its introduction as a so- affect the sophistication of the algorithms phisticated Tamagotchi. In fact, academic in- available on board. Although the Lego RCX’s terest has helped guide Sony’s own choice of Hitachi H8 microcontroller lists at 16 mega- next-generation AIBO features and software hertz and 32 kilobytes of memory, the over- support. As for two-legged platforms, the Uni- head of the firmware and interpreter yield versity of Freiburg has already prototyped a about 10 kilobytes and 500 hertz throughput soccer team of Robosapiens running from for a typical user—slightly better with alterna- handheld computers.5 tive versions of the firmware (Gockley 2003). Similarly, the software that supports AI ro- The Handy Board’s Motorola 2 megahertz botics education will continue to mature. Im- 68HC11 is not too different, though its Interac- provements and software resources come from tive C interface additionally provides for up to at least three directions. First, commercial 4 threads. The Game Boy’s 32-bit 16 megahertz third-party products such as the multiplatform ARM processor, however, does sport noticeably Webots simulator and the RoboJDE develop- deeper computational pockets: 256 kilobytes of ment environment (Michel 2004) offer turnkey memory, 92 kilobytes of video random-access interfacing and development capabilities. A memory, and up to 32 threads. The AIBO’s cur- second source of software is the research com- rent 576-megahertz 64-bit RISC chip pushes munity: the Stage and Gazebo simulators performance to yet another level—particularly demonstrate the benefits of developing a re- important because vision is its primary sensor. search-inspired codebase into a more general- Although additional capabilities certainly ease purpose tool. Finally, enthusiasts of all stripes students’ investigation of a wide variety of al- donate time and effort to make the materials gorithms, the limitations of an RCX or Handy they have developed available to the commu- Board can also serve to motivate research is- nity. Each of the articles in this AI Magazine sues, for example, in sensor-limited robotics special issue contributes to this effort by en- and resource-bounded reasoning. couraging educational access to software and hardware that AI researchers use as a matter of Onward course. The available resources for incorporating ro- Figures 1, 2, and 3 list a number of factors bots into AI education are considerable, yet that inform whether and how physical agents they may only hint at the opportunities on the fit into an AI course or sequence. Yet the figures horizon. One emerging possibility is that some do not address the most important pedagogical educational robots will blur away from the sta- factor in choosing a platform: the AI topics tus quo of “complete systems” into peripheral supported or enhanced by each resource. The form factors that use existing cameras, laptop next section outlines the variety of options and computers, and networking capabilities. By re- approaches available for undergraduate curric- lying on mobile computational devices such as ula in AI robotics. game consoles and handheld computers, the Palm Pilot robot kit (PPRK), XPort Robot Con- AI Robotics Curricula troller, and XBC have begun this process, as have trial laptop-based systems like the Evolu- Robot platforms can be introduced into AI ed- tion ER1. Leveraging existing infrastructure of- ucation in a variety of ways, ranging from fers an opportunity for producing physical adding a single robot assignment to an AI agents with high-end capabilities at a lower course to designing a complete AI robotics cost than today’s kit-based systems. Most visi- course to adding AI material to an integrated bly represented by iRobot’s vacuum cleaners,3 robot engineering course. The early adopters of the nascent home robot industry similarly low-cost robotics in AI education began by us- promises to make autonomous mobile compu- ing robot platforms to teach behavior-based or

16 AI MAGAZINE Articles

IR sensors There are two fundamentally different IR sensors: detectors that report a bit indicating the presence or $10-40 absence of an object, and rangers that return a value proportional to the object’s distance. Ranges from Sensor range: 4 - 80 cm 4 cm to 30 cm or 10 cm to 80 cm are widely available.

Sonar sensors Sonar sensors convert the time-of-flight of an ultrasonic ping into range. Sonar and IR sensors $34-60 enable the building of evidence-grid represen- tations of the environment when encoders/ rotation sensors are available (Martin and Sensor range: 4 - 1000 cm Moravec 1996).

Cameras The CMUCam2, and to a lesser extent the original CMUcam, can provide raw images and, more $109-200 usefully, statistics about sequence-tracked regions of uniform color. They do not interface with the Sensor resolution: RCX, however (Rowe et al. 2002). 176 x 255 pixels

Figure 3. Three of the Most Commonly Used Sensors beyond Basic Light and Touch Feedback for Lego Mindstorms, the Handy Board, Xport, or PPRK Controllers.

reactive architectures (Martin 1996). This was a situation through trigger events. The next driven in large part by the perceived limitations level builds representations such as maps for of the processing capabilities of the platforms navigation. A subsequent tier interprets infor- (for example, Murphy 2000). Work on using mation at the knowledge level for reasoning these platforms that goes beyond reactive con- and planning. This view can provide a natural trol structures has flourished since then. approach for incorporating robotics within the Though necessarily incomplete, table 1 shows a various subfields of AI. Alternatively this view sampling of the wide range of AI topics that can provide a means of exploring a progression have been successfully implemented on low- of AI topics by first beginning with developing cost robot platforms, along with the platforms a deliberative robotic system to navigate and used and links to accompanying curricular ma- solve a task, then progressing to behavior-based terial. control, and finally developing hybrid ap- Table 1 omits an important consideration for proaches that lead to complete solutions. AI educators contemplating the use of robots: The choice of platform directly affects the how well they fit into the overall syllabus of construction of an AI curriculum using physi- course topics. One approach presents AI robot- cally embodied agents. Robot kits such as the ics as a separate topic, as in Russell and Norvig Lego RCX, Handy Board, or Xport use Lego (2003). In his article in this issue, Frederic building blocks and gears to provide an easy Crabbe offers an alternative view that deeply means for students to build mechanical struc- integrates AI and robotics as a process of ab- tures. While students tend to be drawn to the straction and interpretation. This framework creative and experimental nature of building consists of layers of abstraction ranging from their own robot, the use of such platforms must signal processing to long-term behaviors. be weighed against the time students spend Crabbe shows how different approaches to building and rebuilding. Alternatives ap- teaching robotics, including teaching AI using proaches provide prebuilt platforms (for exam- robotics, can be formulated using this frame- ple, Bruder and Wedeward [2003]) or plans for work. This framework describes subfields using and examples of platforms to aid student con- varying levels of data abstraction and interpre- struction. These choices affect the student’s ex- tation. At the level of behavior-based robotics, perience. For example an ordinary differential data is interpreted as the presence or absence of drive (a motor driving one wheel on each side)

SPRING 2006 17 Articles

will undergo significant rotational drift due to learned knowledge to real-world application, natural differences in motors. In contrast, a dri- but suffer from several challenges. The first ve using a differential or dual differential, in challenge is structural: how to both reinforce which a single motor drives both wheels while the learned knowledge and ensure that it ex- a second motor provides turning, will provide cites and motivates students. An instructor more accurate straight-line motion and turning must balance between open-ended competi- (Mayer 2004). tion problems that may cause stress and more Are robots worth the effort required to incor- routine, step-by-step competitions that may porate them in an undergraduate AI curricu- not stimulate students intellectually. A second lum? Existing resources have a major impact challenge involves supporting the cross-disci- on effort and educational effectiveness. For plinary contributions of most robotic laborato- some platforms, user communities and interest ry exercises, while still developing a skill set rel- groups provide ready-made educational re- evant to the overarching course. As a means of sources via books (such as Martin [2000]), dis- addressing these two challenges, low-cost ro- cussion boards and software (for example, the botic platforms allow students to work with Lego User’s Group),6 and commercial sites systems of sufficient complexity to engage (such as HiTechnic Products).7 Table 1 offers them at an appropriate level; instructor-provid- pointers to some of these resources—particular- ed abstractions ease the difficulty as appropri- ly those of use to educators designing under- ate. Low-cost robot platforms further allow a graduate AI curricula. multidisciplinary approach to AI education in Curriculum, however, is only part of the sto- which diverse student learning styles are val- ry. Robots have also spawned vibrant commu- ued and expanded. Successful variations in lab- nities of educators and researchers who share oratory organization have included segment- their work at competitions and exhibitions. ing the class into teams of different majors, as The next section looks at the impact of some of well as pairing together novice and beginning these forums. students into one unified group (Weinberg et al. 2005). Local competitions ultimately provide stu- Robot Contests dents with a goal to strive toward. Whether Robot contests provide a forum in which stu- through the use of laboratory exercises or dents design and build robots to solve a specific through design teams with a faculty advisor, lo- engineering problem. Competitions represent cal competitions provide a safe atmosphere in the integration of many facets of engineering which to stretch the creativity of the student, and science—from mechanical construction to while validating the theory behind robotic us- computer programming. They are excellent op- age. What’s more, a capstone experience need portunities to reinforce the relationship math not require competition between student and science have on tangible real-world appli- teams. Student groups may instead aim toward cations. Pedagogically, competitions can be use- meeting a set of prespecified criteria (Parsons ful in motivating and developing the social as- and Sklar 2004) or in the creation of a creative pects of teamwork and collaboration, and they exhibition (Turbak and Berg 2002). can be effective in bridging the gap between cursory and deep mastery of subject matter. National Competitions A number of robot contest traditions, such as The last five years have seen a number of wide- RoboCup and the AAAI , reaching robotics competitions grow in scope. have emerged as natural motivators to students Three of the most influential are RoboCup, and researchers alike. These opportunities can AAAI’s Competition, and Bot- excite students about engineering and science. Ball. All of these competitions integrate several Even so, special care must be taken to link the types of engineering and science to solve spe- experience with a mathematics and science cific problems within a given domain. These foundation. If not, robotics competitions be- different domains, such as mechanical struc- come a tool only to encourage, but not to ture, software, and electronics, must be fused teach, leaving a formidable gap between enter- into a common platform to create a functional tainment and education. system for diverse problem sets. There has been considerable recent concern with the large Local Competitions amount of effort required of the students for Local competitions are a part of many robotics competition—especially at the national level. courses; often they are the culmination of a se- To complement this trend, aspects of outreach, quence of laboratory exercises for building and coupled with the competition environment, programming a robot. Such labs effectively link have begun to permeate the community

18 AI MAGAZINE Articles

Topics Platform Knowledge representation Lego RCX with off-robot processing Heuristicsearch Resource: www.csc.villanova.edu/~klassner Landmark navigation Multirobot communication Probabilistic localization (particlefiltering) Evolution ER-1 Machine vision Resource: www.cs.hmc.edu/~dodds/courses/ Planning Lego RCX, Handy Board Wave-front navigation Resource: roboti.cs.siue.edu/ Hybridcontrol Neural and Bayesian networks Handy Boardwith on- and off-robot processing; Lego RCX Probabilisticplanning Resource: www.cs.hmc.edu/roboteducation/ Vector field histogram mapping and navigation Probabilistic localization (particlefiltering) Neural networks Pyrosimulator Computer vision Resource: emergent.brynmawr.edu/~dblank/pyro/ Geneticalgorithms Uncertainty Lego RCX Planning and control Resource: www.cs.brown.edu/courses/cs148/ Additional Resources Layersofabstraction, kinematics, behavior-based control, Byo-bots, RobixRascal (roboticarm), RugWarrior dead reckoning Resource: www.cs.usna.edu/~crabbe/teaching.html Reactive and behavior-based control Lego Mindstorms, eLeague robot soccer Resource: www.cs.columbia.edu/~sklar/ Behavior-based control,planning, wave-front navigation, Lego Mindstorms, Handy Board, Xport sensors, kinematics Resource: roboti.cs.siue.edu/ Behavior-based control,neural networks Pyrosimulator Resource: www.cs.uml.edu/~holly/#teaching Behavior-based control,planning, computer vision, Sony AIBO multirobot communication Resource: www-2.cs.cmu.edu/~mmv/ Behavior-based control,navigation, multirobot Sony AIBO, Lego Mindstorms, Byo-bots, Gamebots interaction Resource: www.cs.uno.edu/~sheila/ Behavior-based control PPRK, Handy Board, Lego Mindstorms Resource: www.personal.psu.edu/faculty/r/l/rla5/ Integrated engineering Handy Board Resource: www.generalrobotics.org Sensors Lego Mindstorms, Handy Board Resource: www.philohome.com/sensors.htm Message board Lego Mindstorms, Handy Board Resource: news.lugnet.com/robotics/

Table 1. A Sampling of AI Topics Implemented on Low-Cost Robot Platforms. Top: A sampling of advanced AI topics taught using low-cost robotics with URLs. Bottom: Additional resources of use to educators who incorporate robots within an AI course.

(Baltes, Sklar, and Anderson 2004; Miller and mer Robotic Camp (Nourbakhsh at al. 2005) Winton 2004; Stein, Schein, and Miller 2002). and NASA’s Athena Student Interns Program National outreach efforts complement robot- have engaged groups of students in hands-on ics competitions through mentoring, open- activities that reinforce math and science skills source web curriculum, as well as other tradi- through robotic tasks. Colleges with a history of tional means of outreach. Specific programs, competition have begun to establish programs such as the Carnegie Mellon University Sum- to mentor less experienced teams. Outreach al-

SPRING 2006 19 Articles

so focuses on providing robotics learning tools son’s set of products and publications. To pro- to students and educators. On-line web-based vide a jumping-off point for further investiga- services such as NASA’s Robotics Education Pro- tion of the available platforms, topics, and con- ject and Imagiverse Robotics have disseminated tests, we have a web page that offers an online lesson plans, interviews, and classroom activi- compendium of the information in this ties. Whether competition- or outreach-based, article.8 We hope that this issue’s overview of these programs support the same goals: to mo- robot resources will serve as an invitation for AI tivate and encourage students through physi- educators of all backgrounds to reflect on their cally embodied computation. AI classes and the role of physically embodied agents in them. Perspective Notes This overview of —plat- 1. www.ridgesoft.com. forms, curricula, and contests—sets the stage 2. www.kipr.org. for the five papers that follow. Each one inves- 3. www.irobot.com. tigates a facet of the robotics resources that can 4. www.roombacommunity.com. support undergraduate AI education. Each, too, 5. www.nimbro.net/rs. emphasizes the considerable common ground 6. www.lugnet.com. between the fields of “artificial intelligence” 7. www.hitechnic.com. and “robotics.” Although those terms can be 8. www.cs.hmc.edu/roboteducation. used to connote distinct fields of research, it is often the overlap between the two that moti- vates undergraduate interest in each. To charac- References terize this synergy, Frederic Crabbe describes a Arkin, R. 2000. Autonomous Robotics Education at framework in this issue of AI Magazine in which Georgia Tech. IEEE Intelligent Systems 15(6): 15. AI robotics serves as a unifying theme for a Avanzato, R. 2004. Interfacing Handheld Computers to broad spectrum of topics in both of these fields. Mobile Robots. In Accessible Hands-on Artificial Intelli- The educational resources available for AI ro- gence and Robotics Education, ed. L. Greenwald, Z. botics also highlight these fields’ common Dodds, A. Howard, S. Tejada, and J. Weinberg, 68–71. Technical Report SS-04-01. Menlo Park, CA: AAAI Press. ground. Frank Klassner’s article brings a ubiqui- tous AI tool, Lisp, to the Lego Mindstorms plat- Bagnall, B. 2002. Core Lego Mindstorms Programming. Englewood Cliffs, NJ: Prentice-Hall PTR. form, facilitating the incorporation of robotics into existing curricula. The Pyro project by Baltes, J.; Sklar, E.; and Anderson, J. 2004. Teaching with RoboCup. In Accessible Hands-on Artificial Intel- Douglas Blank, Deepak Kumar, Lisa Meeden, ligence and Robotics Education, ed. L. Greenwald, Z. and Holly Yanco uses Python to provide a Dodds, A. Howard, S. Tejada, and J. Weinberg, widely scalable abstraction for teaching AI 131–135. Technical Report SS-04-01. Menlo Park, CA: across a variety of hardware and software plat- AAAI Press. forms. Both of these articles provide concrete Bishop, B.; Piepmeier, J.; Piper, G.; Knowles, K.; Ho, starting points for smoothly integrating real or K.; and Husock, B. 2004. The Use of Low-Cost RC Ser- simulated physical agents into an AI classroom. vos in a Robotics Curriculum. In Accessible Hands-on Physical agents can support existing AI cur- and Robotics Education, ed. L. ricula; they, too, can motivate topics relatively Greenwald, Z. Dodds, A. Howard, S. Tejada, and J. new to undergraduate AI. Lloyd Greenwald, Weinberg, 52–56. Technical Report SS-04-01. Menlo Donovan Artz, Yogi Mehta, and Babak Shirmo- Park, CA: AAAI Press. hammadi outline a curriculum in which stu- Bruder, S., Wedeward, K. 2003. Robotics in the Class- dents implement and investigate probabilistic room. IEEE Robotics & Magazine 10(2): spatial reasoning and machine learning algo- 25–29. rithms using the inexpensive Lego Mindstorms Challinger, J. 2005. Efficient Use of Robots in the Un- and Handy Board platforms. The article by dergraduate Curriculum. In Proceedings of the Thirty- Sixth SIGCSE Technical Symposium on Computer Science Manuela Veloso, Paul Rybski, Scott Lenser, So- Education, 441–445. New York: Association for Com- nia Chernova, and Douglas Vail spotlights the puting Machinery. capabilities of Sony’s AIBO robotic dog as a Coradeschi, S.; and Malec, J. 1999. How to Make a pedagogical tool that bridges with RoboCup Challenging AI Course Enjoyable Using the RoboCup competitions and a vibrant research communi- Soccer Simulation System, 120–124. In RoboCup-98: ty. Robot Soccer World Cup II: Lecture Notes in Artificial In- This introduction necessarily falls short of a telligence, vol. 1604, ed. M. Asada and H. Kitano. complete accounting of available resources for Berlin: Springer. teaching AI robotics to undergraduates. Any Crabbe, R. 2004. Unifying Undergraduate Artificial such list would quickly succumb to next sea- Intelligence Robotics: Layers of Abstraction over Two

20 AI MAGAZINE Articles

Channels. In Accessible Hands-on Artificial Intelligence Notions of Control in Undergraduates Who Design and Robotics Education, ed. L. Greenwald, Z. Dodds, A. Robots. In Constructionism in Practice: Designing, Howard, S. Tejada, and J. Weinberg, 2–7. Technical Thinking, and Learning in a Digital World, ed. Y. Kafai Report SS-04-01. Menlo Park, CA: AAAI Press. and M. Resnick. Mahwah, NJ: Lawrence Erlbaum As- Danyluk, A. 2004. Using Robotics to Motivate Learn- sociates. ing in an AI Course for Non-Majors. In Accessible Martin, F. 2000. Robotic Explorations: An Introduction Hands-on Artificial Intelligence and Robotics Educa- to Engineering through Design. Englewood Cliffs, NJ: tion, ed. L. Greenwald, Z. Dodds, A. Howard, S. Teja- Prentice-Hall. da, and J. Weinberg, 93–96. Technical Report SS-04- Martin, F., and Pantazopoulos, G. 2004. Designing 01. Menlo Park, California: AAAI Press. the Next-Generation Handy Board. In Accessible Gockley, R. 2003. Multi-Robot Communication with Hands-on Artificial Intelligence and Robotics Education, Lego Mindstorms Web Site. Complexity Problems ed. L. Greenwald, Z. Dodds, A. Howard, S. Tejada, and Page. Pittsburgh, PA: Carnegie Mellon University. J. Weinberg, 77–81. Technical Report SS-04-01. Men- Visited 1 April 2005: www.contrib.andrew.cmu.edu/ lo Park, CA: AAAI Press. ~rgockley/legos/bad.html). Martin, M., and Moravec, H. 1996. Robot Evidence Greenwald, L., and Artz, D. 2004. Teaching Artificial Grids. Technical Report CMU-RI-TR-96-06. Robotics Intelligence with Low-Cost Robots. In Accessible Institute, Carnegie Mellon University, Pittsburgh, PA. Hands-on Artificial Intelligence and Robotics Educa- Mayer, G. 2004. Implementation of a Deliberative Ro- tion, ed. L. Greenwald, Z. Dodds, A. Howard, S. Teja- bot Control Architecture on an Inexpensive Robot da, and J. Weinberg, 35–41. Technical Report SS-04- Platform, Master’s Thesis, Department of Computer 01. Menlo Park, CA: AAAI Press. Science, Southern Illinois University, Edwardsville, IL. Harlan, R. 2005. Creating Emergent Behaviors: Two Michel, O. 2004. Webots: Professional Mobile Robot Robotics Labs That Combine Reactive Behaviors. In Simulation. International Journal of Advanced Robotic Proceedings of the Thirty-Sixth SIGCSE Technical Sympo- Systems 1(1), 39–42. sium on Computer Science Education, 441–445. New Miller, D., and Winton, C. 2004. Botball Kit for York: Association for Computing Machinery. Teaching Engineering Computing. In Proceedings of Huang, W., and Beevers, K. 2004. Topological Map- the ASEE National Conference in Salt Lake City, UT. ping with Sensing-Limited Robots. In Proceedings of the Washington, DC: American Society of Electrical Engi- Sixth International Workshop on the Algorithmic Founda- neers. tions of Robotics, 367–382. Berlin: Springer-Verlag. Murphy, Robin. 2000. Introduction to AI Robotics. Imberman, S. 2004. A Laboratory Exercise Using Cambridge, MA: MIT Press. LEGO Handy Board Robots to Demonstrate Neural Nourbakhsh, I.; Crowley, K.; Bhave, A.; Hamner, E.; Networks in an Artificial Intelligence Class. In Acces- Hsiu, T.; Perez-Bergquist, A.; Richards, S.; Wilkinson, sible Hands-on Artificial Intelligence and Robotics Educa- K. 2005. The Robot Autonomy Mobile Robotics tion, ed. L. Greenwald, Z. Dodds, A. Howard, S. Teja- Course: Robot Design, Curriculum Design and Edu- da, and J. Weinberg, 77–81. Technical Report cational Assessment. Autonomous Robotics Journal SS-04-01. Menlo Park, CA: AAAI Press. 18(1)(January): 103–127. Klassner, F. 2004. A Tool for Integrating Lisp and Ro- Nuessle, T.; Kleiner, A.; and Brenner, M. 2004. Ap- botics in AI Agents Courses. In Accessible Hands-on proaching Urban Disaster Reality: The ResQ Firesim- Artificial Intelligence and Robotics Education, ed. L. ulator. In RoboCup 2004: Robot Soccer World Cup VIII. Greenwald, Z. Dodds, A. Howard, S. Tejada, and J. Lecture Notes in Computer Science, vol. 3276. Berlin: Weinberg, 24–29. Technical Report SS-04-01. Menlo Springer. Park, CA: AAAI Press. Parsons, S., and Sklar, B. 2004. Teaching AI Using Konolige, K.; Fox, D.; Ortiz, C.; Agno, A.; Eriksen, M.; LEGO Mindstorms. In Accessible Hands-on Artificial Limketkai, B.; Ko, J.; Morisset, B.; Schulz, D.; Stewart, Intelligence and Robotics Education, ed. L. Greenwald, B.; and Vincent, R. 2004. Centibots: Very Large Scale Z. Dodds, A. Howard, S. Tejada, and J. Weinberg, Distributed Robotic Teams. In Proceedings of the 8–13. Technical Report SS-04-01. Menlo Park, CA: Eighth International Symposium of Experimental Robot- AAAI Press. ics. Berlin: Springer. Reshko, G.; Mason, M.; and Nourbakhsh, I. 2002. Kuhnel, C., and Zahnert, K. 1997. Basic Stamp. New Rapid Prototyping of Small Robots. Technical Report York: Elsevier. CMU-RI-TR-02-11. Robotics Institute, Carnegie Mel- Kumar, D., and Meeden, L. 1998. A Robot Laboratory lon University, Pittsburgh, PA. for Teaching Artificial Intelligence. SIGCSE Bulletin Rowe, A.; Rosenberg, C.; and Nourbakhsh, I. 2002. A 30(1): 341–344. Low Cost Embedded Color Vision System. In Proceed- LeGrand, R. 2004, Closed-Loop Motion Control for ings of the International Conference on Intelligent Robots Mobile Robotics. Circuit Cellar 169: 34–47. and Systems, vol. 1, 208–213. Piscataway, NJ: Institute LeGrand, R.; Machulis, K.; Miller, D. P.; Sargent, R.; of Electrical and Electronics Engineers. and Wright, A. 2005. The XBC: A Modern Low-Cost Russell, S., and Norvig, P. 2003. Artificial Intelligence: Mobile Robot Controller. In Proceedings of the A Modern Approach, 2nd ed. Englewood Cliffs, NJ: IEEE/RSJ International Conference on Intelligent Robots Prentice-Hall. and Systems (IROS 2005). Piscataway, NJ: Institute of Stein, C.; Schein, D.; and Miller, D. P. 2002. AAAI Electrical and Electronics Engineers. Hosts the National Botball Tournament! AI Magazine Martin, F. 1996. Ideal and Real Systems: A Study of 23(1): 51–54.

SPRING 2006 21 Articles

Zachary Dodds is an associate professor of computer science at Harvey Mudd College who has taught hands-on, AI-based com- puter vision and robotics for the past six years. His interests include vision-based robot mapping and developing hardware and software to help make low-cost robots more accessible within the CS curriculum. Zach is on sab- batical at Carnegie Mellon University in 2005–06. He AAAI Mobile Robot is reachable at [email protected]. Competition and Exhibition Lloyd Greenwald is currently a member of technical staff at Bell The Fifteenth Annual Robot Competition and Exhibition will be Labs in the Internet Research De- held in Boston, MA, from July 16–20, 2006, in conjunction with partment pursuing research on ad- the Twenty-First National Conference on Artificial Intelligence. vanced algorithms for network se- curity including anomaly We invite your participation in this exciting competition, which detection, vulnerability analysis, will feature a scavenger hunt, and open interaction task and ro- penetration testing, wireless secu- bot challenge, and a robot exhibition, as well as a workshop that rity, and mobile ad hoc networks. takes place on the last day of the conference. Greenwald received his Ph.D. and Sc.M. in computer science from Brown University and a B.S.E. in com- For details visit www.aaai.org/Conferences/AAAI/2006/aaai06ro- puter science and engineering from the University of bots.php or palantir.swarthmore.edu/aaai06/ Pennsylvania. His robotics education contributions were primarily pursued while he was an assistant pro- Hurry! The deadline for participation is fessor of computer science and director of the Intelli- May 15th. gent Time-Critical Systems Lab at Drexel University. He can be reached at [email protected].

Jerry B. Weinberg is an associate professor and chair of the Com- puter Science Department at Stone, P. 2004. RoboCup as an Introduction to CS Re- Southern Illinois University Ed- search. In RoboCup 2003: Robot Soccer World Cup VII. wardsville. He teaches courses and Lecture Notes in Computer Science, vol. 3020, ed. D. conducts research in AI, robotics, Polani, B. Browning, A. Bonarini, and K. Yoshida, and HCI. Weinberg formed the Ro- 284–295. Berlin: Springer. botics Project Group, which has introduced robotics projects in Sutherland, K. 2000. Undergraduate Robotics on a various computer science and engineering courses, Shoestring. IEEE Intelligent Systems 15(6): 28–31. and has initiated various robotics outreach programs. Turbak, F., and Berg, R. 2002. Robotic Design Studio: He has also mentored both high school and college Exploring the Big Ideas of Engineering in a Liberal student teams for regional and national robot com- Arts Environment. Journal of Science Education and petitions. Information about robotics activities at Technology 11(3)(September): 237–253. SIUE can be found at http://roboti.cs.siue.edu. Vaughan, R.; Gerkey, B.; and Howard, A. 2003. On Device Abstractions for Portable, Reusable Robot Ayanna Howard is an associate Code. In Proceedings of the IEEE/RSJ International Con- professor in the School of Electri- ference on Intelligent Robots and Systems (IROS 2003), cal and Computer Engineering at 2121–2427. Piscataway, NJ: Institute of Electrical and the Georgia Institute of Technolo- Electronics Engineers. gy. She received her B.S. from Veloso, M.; Lenser, S.; Vail, D.; Rybski, P.; Aiwazian, Brown University, her M.S.E.E. N.; and Chernova, S. 2004. CMRoboBits: Creating an from the University of Southern California, and her Ph.D. in elec- Intelligent AIBO Robot. In Accessible Hands-on Artifi- trical engineering from the Uni- cial Intelligence and Robotics Education, ed. L. Green- versity of Southern California in 1999. Her area of re- wald, Z. Dodds, A. Howard, S. Tejada, and J. Wein- search is centered on the concept of humanized berg, 57-62. Technical Report SS-04-01. Menlo Park, intelligence, the process of embedding human cogni- California: AAAI Press. tive capability into the control path of robotic sys- Weinberg, J.; White, W.; Karacal, C.; Engel, G.; and tems. She can be reached at [email protected] Hu, A. 2005. Multidisciplinary Teamwork in a Robot- ech.edu. ics Course. In Proceedings of the Thirty-Sixth SIGCSE Technical Symposium on Computer Science Education Sheila Tejada is affiliated with Tulane University. (SIGCSE 2005), 446–450. New York: ACM Press.

22 AI MAGAZINE