AI Magazine Volume 27 Number 1 (2006) (© AAAI) Articles Using Educational Robotics to Motivate Complete AI Solutions Lloyd Greenwald, Donovan Artz, Yogi Mehta, and Babak Shirmohammadi ■ Robotics is a remarkable domain that may be suc- (Nilsson1984) to more recent work on multia- cessfully employed in the classroom both to moti- gent systems and machine learning in vate students to tackle hard AI topics and to pro- RoboCup (Stone 2000). vide students experience applying AI represen- In addition to providing inspiration, explor- tations and algorithms to real-world problems. ing artificial intelligence representations and This article uses two example robotics problems to illustrate these themes. We show how the robot ob- algorithms using robotics helps students to stacle-detection problem can motivate learning learn complete solutions. A complete solution neural networks and Bayesian networks. We also is one in which a student considers all the de- show how the robot-localization problem can mo- tails of implementing AI algorithms in a real- tivate learning how to build complete solutions world environment. These details range from based on particle filtering. Since these lessons can system design, to algorithm selection and im- be replicated on many low-cost robot platforms plementation, to behavior analysis and experi- they are accessible to a broad population of AI stu- mentation, to making the solution robust in dents. We hope that by outlining our educational the face of uncertainty. In our classes we find exercises and providing pointers to additional re- sources we can help reduce the effort expended by that robotics problems encourage students to other educators. We believe that expanding hands- investigate how AI algorithms interact with on active learning to additional AI classrooms pro- each other, with non-AI solutions, and with a vides value both to the students and to the future real-world environment. Students investigate of the field itself. how to convert sensor data into internal data structures, how to weigh the costs and benefits of physical exploration, whether or not to use rtificial intelligence has a number of offline simulation and tools, and how to deal wonderful domains that help motivate with the severe resource limitations and time Aundergraduate students to enroll in AI constraints of embedded computation. Despite courses and work hard to grasp advanced AI the added costs of building complete solutions, representations and algorithms. One particu- experience with real-world environments helps larly compelling domain is robotics. Robotics ground lessons and stimulates thinking about combines the fantasy of science fiction with new challenges and solutions. practical real-world applications and engages Although robotics has inspired research on both the imaginative and sensible sides of stu- AI representations and algorithms, until re- dents. Examples of robotics-inspired AI span cently the cost and size of these platforms has from early work on AI planning with Shakey made it difficult to convey the same line of in- Copyright © 2006, American Association for Artificial Intelligence. All rights reserved. 0738-4602-2006 / $2.00 SPRING 2006 83 Articles vestigation to undergraduate students. Low- neural networks or Bayesian networks and im- cost robot platforms, as surveyed earlier in this plementing the solution on a Handy Board– issue (Dodds et al. 2006), have the potential to based robot. The second theme is presented in mitigate this cost barrier. Investigators have be- the context of solving a robot-localization gun to report successful demonstrations of problem using particle filtering and imple- hands-on robotics using low-cost platforms in menting the solution on an RCX-based robot. AI courses (Greenwald and Kopena 2003; May- The projects described in this article were de- er, Weinberg, and Yu 2004, Dodds et al. 2004). signed primarily as part of a stand-alone robot- These reports, as well as the instructional mate- ics course for computer science majors. The rial on advanced AI topics in this special issue, hands-on topics covered in this course include provide critical supplements to existing texts robot structure, drive trains, sensors (infrared, and curricular material on general undergradu- photo, bump, active sonar), control (open ate artificial intelligence, engineering, and loop, closed loop), simple behaviors (obstacle computer science topics. detection, light following, obstacle avoidance), Even at low cost ($300–$600 per robot kit), it reactive control, multiple behavior coordina- is possible to teach advanced AI with the hard- tion (algorithmic, priority-based, subsump- ware and software resources of these platforms. tion), odometry using wheel encoders, forward Low-cost platforms provide varying levels of kinematics and localization, inverse kinemat- mobility, sensors, development environments, ics, way-point navigation, simple feature detec- and processing power. Lego-based platforms tion using sonar, probabilistic localization, lo- like the RCX (Klassner 2002), or Handy Board cal and global map building, sensor fusion, (Martin 1999) are designed for flexible experi- vector field histograms, and simple path plan- mentation with mobility. Development envi- ning. For interested students this course pro- ronments include many standard program- vides hands-on experience that complements ming languages like Java, C, Pascal, Forth, and material taught in artificial intelligence courses ADA. Even Lisp is available, as detailed else- for advanced undergraduate and early graduate where in this issue of AI Magazine (Klassner students. The projects in this article have also 2006). Commonly available sensors include been used in workshops, including a recent photoresistors, infrared, touch, wheel encoders NSF-sponsored workshop for teachers learning (for odometry), and, most impressively, sonar. to use the RCX in education. Please refer to Greenwald and Kopena (2003) and Greenwald With respect to implementing advanced AI al- and Artz (2004) for further information on the gorithms, the processing power of low-cost structure and labs of this course. platforms can be problematic. At the low end, we get from 2 megahertz to 16 megahertz of central processing unit (CPU) speed and around 32 kilobytes of random-access memory Obstacle Detection with Neural (RAM). The education lessons discussed in this Networks, Bayesian Networks, article have been carefully designed to work successfully with such constrained resources. and the Handy Board We have found that these resource constraints A typical task in a robotics class is to program a often provide additional educational opportu- robot to detect and avoid obstacles. In our ro- nities to investigate AI topics like resource- bot-building lab course (Greenwald and Kope- bounded reasoning. These platforms addition- na 2003) we initially assigned this as a non-AI ally provide communications protocols so that exercise in processing noisy sensor readings. off-board processing can be included in com- We chose to use inexpensive and unreliable in- plete AI solutions. frared sensors as proximity detectors for this We have used low-cost robot platforms in exercise. In lectures we taught students stan- the classroom to teach many AI topics includ- dard methods for processing sensor data, in- ing heuristic search, planning, probabilistic cluding ambient light correction, surface re- planning, representing uncertainty, and ma- flectance calibration, thresholding, averaging, chine learning. In this article, we outline our and hysteresis. The students were later asked to classroom-tested material for teaching Bayesian use these processing methods to classify sensor networks, neural networks, and particle filter- data into one of four obstacle position states: ing. This article is organized around the two ed- left, right, center, or none. ucational themes just introduced: (1) robotics This seemingly simple exercise repeatedly problems motivate AI solutions, and (2) robotics led to frustration and failure. Students reported problems encourage complete AI solutions. The frequent invalid infrared readings in any non- first theme is presented in the context of solv- shielded application, indicating that these sen- ing a robot obstacle-detection problem using sors were useless as proximity detectors for ob- 84 AI MAGAZINE Articles stacle detection and avoidance. The following mitter while detecting the ambient light being exercise helped us determine whether the received by the IR sensor’s receiver. source of the problem was the sensors or the The IR sensor transmits infrared light away programmers, and at the same time uncovered from the robot. Reflected IR signals are received our first educational theme: robotics problems if an object is sufficiently near the sensor. The motivate AI solutions. color, surface texture, angle, and other factors We initiated a project to see whether or not affect the distance required to register reflected these sensors were actually useless. The result- IR signals in the IR sensor’s receiver. High ing project demonstrated not only that these amounts of reflected infrared light yield high inexpensive sensors could be used for obstacle signal values. If there is little or no reflected in- detection but also that their inherent unrelia- frared, the IR sensor’s receiver registers a low bility provides a practical motivation for teach- signal value. ing advanced artificial intelligence techniques The sensors are placed approximately in a for sensor processing.