Understanding Behavior 1 Running head: UNDERSTANDING BEHAVIOR FROM THE GROUND UP Understanding Behavior from the Ground Up: Constructing Robots to Reveal Simple Mechanisms Underlying Complex Behavior

Students often have difficulty getting past the use of folk-psychological terms (e.g. wants, loves, fears) when explaining behavior. They assume that complex, smart behaviors require similarly complex causal structures in the head. For Introductory Psychology, we developed a two-week robotics project to show students that complex behaviors can emerge from simple mechanisms. The project combined lectures, demonstrations of simple robots navigating the classroom, and hands-on robot building and observing experience. Student evaluations showed that they enjoyed the project and learned from their experience. The project accomplished four goals: (1) to engage students in generating and testing hypotheses about robot behavior, (2) to demonstrate the power of a mechanistic approach, (3) to show how social behavior can arise from the simple behaviors of individuals, and (4) to illustrate how organisms change under the pressure of natural selection. Understanding Behavior 3 Understanding Behavior from the Ground Up: Constructing Robots to Reveal Simple Mechanisms Underlying Complex Behavior "A safe rule on encountering any abstract psychological noun is to make it concrete by changing it into the corresponding verb or adverb. Much difficulty and unnecessary controversy can thus be avoided." (Woodworth, 1940, p.19) Imagine two little creatures, bugs perhaps, that are attracted by a light source. When each is far from the source, they race toward the light. However, they differ considerably in their personalities. The first bug approaches the light straight on, slowing, coming to a rest and staying close to the light. The second bug exhibits more erratic behavior, always moving, fluttering slightly away and then back to the light. Sometimes, for no obvious reason, this second bug abandons the light for some distant glimmer. If these bugs were males and the lights were potential mates, one might characterize the two behavior patterns as reflecting true love (or monogamy) versus promiscuity. These words -love, promiscuity -are the kinds of words that people commonly use when trying to “explain” behavior. Students have difficulty getting past such folk-psychological terms. Indeed, when asked to describe even the motions of simple geometric shapes, undergraduates assigned human intentions to the shapes’ actions (Heider & Simmel, 1944). The two bugs and their behaviors, however, are easily generated with components available at a hobby shop – light sensors that drive motors that drive wheels. They were drawn from Braitenberg’s (1984) mini-classic, Vehicles: Experiments in Synthetic Psychology . The “sensory-motor” connections needed to create the two behavioral patterns are illustrated in Figure 1. Notice that Understanding Behavior 4 “monogamy” and “promiscuity” are not components of this mechanistic explanation. We believe these bugs – or their hobby-shop incarnations -offer a way to help students get past the folk-psychological terms to underlying mechanism. Building on Braitenberg, we had students think about and construct sensory-motor devices to mimic complex behavior. Our teaching approach is thus based on current experimental and theoretical investigations in psychology that use robots as demonstrative proofs of proposed mechanisms (Duchon, Warren, & Kaebling, 1988; also see reviews in Arkin, 1996; Hendriks-Jansen, 1996; Pfeifer & Scheier, 1999). --------------------------------Insert Figure 1 about here -------------------------------We developed a two-week robotics project for an introductory psychology course to encourage students to think mechanistically about psychological phenomena and to provide concrete examples of how seemingly complex behaviors might emerge from simple mechanisms. The project was designed to actively engage students and combined demonstrations, lectures, and hands-on building and observing experiences. Our goals were to (1) immerse students in the selfcorrecting process of science by having them generate and test hypotheses, (2) demonstrate the power of a mechanistic approach in explaining complex behavior, (3) show how social behavior can arise from the simple behaviors of individuals, and (4) illustrate how natural selection operates at the level of the whole organism. Students wrote brief daily journal entries on assigned topics, made lab Understanding Behavior 5 observations, and wrote a final (2-3 page) essay on the project (see Appendix A for descriptions of assignments). Method Overview The introductory psychology class had 46 students and was a 4 hour course (2 hours lecture, 2 hours lab). Before the robotics project, the class had covered history of psychology, methods, and adaptation and learning. The project was followed by a section on neuroscience. The robot project comprised one of three 100-point labs, and was worth 14% of the course grade. We divided the two-week project into two sections, the first introducing the mechanistic approach and the second on mechanistic explanations of social behavior and natural selection (Table 1). If desired, the project can be scaled down to only the first week and still retain the basic lessons of the mechanistic approach. --------------------------------Insert Table 1 about here -------------------------------Apparatus We used four Lego Mindstorms Robotics Invention System kits (version 1.0) to build our machines. A Mindstorms kit costs $200.00 and is widely available (see www.legomindstorms.com ). Each kit contained touch and light sensors, motors, wiring, lots of Lego pieces, and a small on-board computer. Also included were illustrated guides for constructing and programming robots (see Appendix B for programming details). The on-board computer had three inputs for sensors, three outputs for motors, and could hold several programs. All pieces of Understanding Behavior 6 the robot kits, including the wiring, could be connected in the familiar snap-together Lego style, so the students did not need specialized engineering skills. To keep the robots in a controlled, observable area during the robot demonstration and labs, we created an arena in the middle of the classroom (Figure 2). --------------------------------Insert Figure 2 about here -------------------------------Procedure Initial Observations of Robot Behavior. We began the first lecture with a live demonstration of SAM (Slightly Autonomous Machine), a small robot with the capacity to seek out light sources while avoiding obstacles in its path (Figure 3). A photographic spotlight provided a bright patch on the floor, and several boxes served as obstacles. As SAM meandered about the classroom, the students (who had not been told anything about SAM’s capacities) vocally labeled “what he was doing”. A list of behaviors was compiled on the chalkboard. Students ascribed intentions and goals to SAM’s behavior, using terms such as “frustration”, “exploring”, “adaptive”, “afraid” (of the dark), “liking”, and “happy”. For their journal assignments, students described mechanisms in SAM’s brain (a small onboard computer) that might be responsible for his behavior. --------------------------------Insert Figure 3 about here -------------------------------During the second lecture we revealed the three sensorimotor rules in SAM’s head, as given in Appendix B. After learning of these simple perceptionUnderstanding Behavior 7 action loops, students discussed why they had given SAM so much psychological power, and considered the possibility that complex behaviors could emerge from simple mechanisms. Constructing Robots. The students divided themselves into four teams, each composed of approximately six “builders” and six “observers” (the students switched roles for the second robot lab). We assigned each team a Mindstorms robot kit and instructed them to build a body capable of navigating an obstacle-filled classroom. Each robot used preprogrammed, identical “brains”, but these programs were not revealed to the students. Students met outside of class in our robot assembly room to construct their machines. To ensure a variety of designs, we encouraged the teams to keep their designs secret. Lab I: The Power of a Mechanistic Approach. The students’ robots were tested in three challenges. Though the robots had identical brains, they had radically different bodies, reflecting the creativity of the teams. We gave observers stopwatches, tape measures, and the job of quantifying the environment (e.g., locations and types of obstacles), robot morphology (e.g., position of sensors relative to the environment and to each other) and robot behavior (e.g., distance traveled, degree of turn, and sounds produced). Once a challenge was announced, observers conferred and decided how to measure their robots’ behavior. Data were collected for each robot over a 5-minute trial. --------------------------------Insert Figure 4 about here -------------------------------Understanding Behavior 8 Challenge 1 required the robots to travel across the classroom floor and use visual landmarks, consisting of thickening black lines on the floor, to avoid a wooden bumper at the far end of the room. The robots had been programmed so that a sudden decrease in activation from the light sensor would initiate a turn. Differences in construction of the bodies and positioning of the sensors led to condsiderable differences in performance. Students rapidly realized that bodies, not just brains, play a strong role in determining a pattern of behavior. Challenge 2 required robots to autonomously navigate an obstacle-laden environment. Observers chose to record how long a robot could “survive” before needing human assistance and frequency of assistance. Some robots appeared quite “smart”, surviving for most of the trial, others quite “dumb”, requiring frequent assistance. They seemed to have quite different brains, but the smart ones differed from the dumb ones only in terms of their mode of locomotion and placement of their touch and light sensors. Challenge 3 showed how preference behavior could be explained mechanistically. Spotlights and speakers were added to the environment to create patches of light and sound (instrumental music). The robots’ programs are illustrated in Table 2 (note that the robots were not sensitive to sound). In addition to the variables recorded in the previous challenges, observers also measured amount of time spent in light versus dark and noisy versus quiet areas. Students debated whether the robots were influenced by sound (they were not). Individual differences were found for survival time and amount of time spent hiding in boxes. Differences in the robots' bodies created variation in patterns of interaction with the environment. The students learned that the robots’ preferences for light versus Understanding Behavior 9 dark, for open areas versus boxes, could be explained without reference to folkpsychological terms. Indeed, the robots’ behavior could be predicted only if those terms were discarded and replaced by a mechanistic framework. Lab II: Natural Selection Acts on the Entire Organism. The second lab consisted of two challenges and demonstrated that evolutionary changes in behavior require forces that act on the entire organism. If nature selects for behaviors that are embodied (determined by both brains and bodies), then selection must involve the whole creature, not some isolated part. Teams built robots to compete in a ping-pong ball “foraging” contest. As before, all “brains” were identical and we did not reveal the program to the students. We told the teams the scoring system for the foraging contest: a point would be awarded for each ball contacted, and an additional point could be gained for each second that the robot remained in contact with a ball. However, two points would be lost each time a robot became ensnared on an obstacle or other robot. Robots began a 10-minute trial with 10 points and success was measured in terms of accumulated points at the end of the trial; running out of points meant death by starvation. The teams again built very different robots; some had scoops or other devices for capturing balls while retaining separate bumpers or antennae for avoiding objects. The contest began with the balls randomly distributed across the classroom floor. Faster robots were generally more successful, but the fastest robot was a failure. Pure speed without appropriate visual and sensory apparatus (such as long feelers for avoiding objects) doomed the robot. Students learned that Understanding Behavior 10 the fitness of a specific behavioral trait (e.g. speed or scoop size) can be accurately assessed only when the behavior of the whole organism is observed. We then shifted the environment and repeated the foraging contest. The robots had to forage for balls in an obstacle-filled environment (Figure 2). In the more complex environment, the slower and more maneuverable robots did better than the faster machines. Students learned that fitness depends not just on an organism’s traits but on the environment that it is in. Simple Mechanisms Can Build Social Behavior. Two lectures focused on mechanistic explanations of social behavior. Group activity that seems complex and directed can be explained, like the behavior of our robots, in terms of interactions between simple sensorimotor rules and a dynamic environment. As examples, we discussed rat pup (Rattus norvegicus) huddling behavior (Alberts, 1978; Schank & Alberts, 1998) and the group anti-predator response of Fall webworms (Hyphantria cunea) (Costa, 1997). The examples catalyzed a lively student discussion of the possibilities of using autonomous robot animals to further investigate the origins of social behavior. The students debated whether their Lego robots could produce patterns of group behavior, and suggested combinations of sensorimotor rules and environmental conditions that would allow their robots to do so. Evaluation and Discussion Students’ understanding of the concepts introduced in the robot project was assessed by the essay and journal assignments (as given in Appendix B). The students’ writing evidenced their understanding of the project, as 43 of the 46 Understanding Behavior 11 students discussed the core concepts of mechanistic explanation and distributed causality. One week after turning in their journals and essays, 42 students filled out a six-item evaluation of the project. The first four questions asked students to rate the project on a 5-point scale for level of difficulty (1 = unreasonable, 5 = too easy), amount of work (1 = unreasonable, 5 = too easy), relevance of lab activities to project lectures (1 = not relevant, 5 = very relevant), and amount learned (1 = nothing, 5 = exceptional amount). Two additional items were short-answer questions asking students to note the best and worst aspects of the project. Students were told that the robot exercise would be repeated in the future, and that their detailed comments and suggestions would be used to help us improve the project. Students indicated that the overall level of difficulty was moderate ( M = 2.86, SD = 0.42; mode = 3.00) and the amount of work also moderate ( M = 2.58, SD = 0.54; mode = 3.00). They found the lab activities to be relevant to the lecture material ( M = 3.79, SD = 0.95; mode = 4.00). Students reported that they learned somewhat more than they expected ( M = 3.55, SD = .94; mode = 4.00). In their written comments, students indicated that the robotics experience had helped them learn about behavior, and that the learning was engaging and fun. They also felt the project could be improved with smaller groups and more time to work with the robots. The students’ increasing sophistication was reflected in their journals. Statements such as “SAM seems to be afraid of light, and he likes to hide behind boxes” were typical of initial journal entries. By the end of the week, the same student quoted above observed: Understanding Behavior 12 Something as simple as turning a sensor plug around can cause a robot that crashes into obstacles to turn away from them....Even with creatures so artificial and simple, it would be useless to study only their minds, because their behaviors are only partially caused by the rules in their heads. This is why behavior must be studied as a property embodied in an organism’s mind, body, and its environment. In their final papers, students showed that they could generalize their ideas from experiments with robots to the workings of nature. One student wrote: Evolution, like Braitenberg’s robot, is governed by simple rules. However, the evolution process or observed “behavior” of nature appears complex because it occurs in a dynamic system. The changing environment, the past history of an organism, the brain of an organism, the current pool of genes in the environment, and random gene mutations, are all factors that influence what is fit and unfit, and what will be selected for in evolution. These factors cause the process of evolution and natural selection to appear so intricate that humans assume it must be controlled by an “intelligent” engineering mind. The robotics project is useful in four ways: (1) It teaches core concepts at the heart of experimental psychology and science. The lessons learned are not restricted to the frontiers of contemporary science. (2) Students learn through “hands-on” and fun experimentation with physical entities. This active-learning approach makes the mechanisms understandable, the rewards of experimentation (that is, correcting false beliefs) fast, and the joy and real-world value of science obvious. (3) The project takes students in one direction in which the world is Understanding Behavior 13 surely heading. Intelligent devices are everywhere -there is no going back. (4) Introductory psychology courses are not populated solely or primarily by students already committed to “science”. The lessons learned from working with the robots -distributed causality, emergent phenomena, mechanistic explanation, and the selfcorrecting nature of science -are useful lessons for everyone. Understanding Behavior 14 ReferencesAlberts, J. R. (1978). Huddling by rat pups: Group behavioralmechanisms of temperature regulation and energy conservation. Journal ofComparative and Physiological Psychology, 92,231-246.Arkin, R. C. (1996). Behavior-Based Robotics.Cambridge, MA: MITPress.Baum, D. (2000). Definitive guide to Lego Mindstorms. Emeryville, CA:Apress.Braitenberg, V. (1984). Vehicles: Experiments in synthetic psychology.Cambridge, MA: MIT Press.Costa, J. T. (1997). Caterpillars as social insects. American Scientist, 85,150-159.Duchon, A. P., Warren, W. H., & Kaelbling, L. P. (1998). Ecologicalrobotics. Adaptive Behavior, 6,473-507.Heider, F., & Simmel, M. (1944). An experimental study of apparentbehavior. American Journal of Psychology, 57,243-259.Hendriks-Jansen, H. (1996). Catching ourselves in the act. Cambridge,MA: MIT Press.In an online salon, scientists sit back and wonder. (1997, December 30).The New York Times,p. 4Schank, J. C., & Alberts, J. R. (1997). Self-organized huddles of rat pupsmodeled by simple rules of individual behavior. Journal of Theoretical Biology,189, 11-25.Woodworth, R. S. (1940). Psychology(4 ed.). New York: Henry Holt. Understanding Behavior 15 Notes1. Michael H. Goldstein is in the Psychology Department and Program in AnimalBehavior. Linda B. Smith is in the Psychology Department and Program inCognitive Science.2. The authors thank John Kruschke, Jennifer Schwade, and three anonymousreviewers for their thoughtful comments on earlier versions of this manuscript.3. Correspondence concerning this article should be addressed to M. H.Goldstein, Indiana University, Psychology Department, 1101 E. 10 St.,Bloomington, Indiana 47405-7007; e-mail: [email protected]. Understanding Behavior 16 Table 1Outline of Robotics Project Day FormatGoalActivity 11 hrlectureIntroduce mechanisticapproach, contrast withfolk-psychologicalapproach.Demonstrate simpleautonomous robot, haveclass label behaviors andgenerate hypotheses. Week 1 21 hrlectureExplore possiblemechanisms, both internaland external, that areresponsible for the robot’sbehavior.Reveal the robot’s simplepre-programmed rules,generate and testhypotheses about causesof robot’s actions. 32 hrlabShow how behavior iscaused by activity atmultiple levels oforganization, emphasizingthe role of the body aswell as the brain.Test autonomousnavigation ability ofstudents’ robots in seriesof increasinglycomplicated environments. 41 hrlectureDemonstrate plausibility ofmechanistic explanationsfor social behavior.Discuss requirements forrobots to be sensitive toeach others’ behavior,thus yielding grouppatterns of interaction. Week 2 51 hrlectureIllustrate utility ofmechanistic approach inexplaining real-worldsocial behavior.Discuss robotic models ofsocial behavior in insects,birds, and rodents. 62 hrlabShow how naturalselection modifies thepopulation of robots.Test students’ robots inseries of survivalchallenges, such as cliffavoidance and efficiencyof foraging for food. Understanding Behavior 17 Figure CaptionsFigure 1.Mechanisms underlying “true love” (vehicle A) and “promiscuous”(vehicle B) behavior. Both vehicles have spatially separated light sensors that drivewheels via negative (inhibitory) connections so that the stronger the light, theslower the approach. The vehicles differ only in whether the connections arestraight or crossed. With straight connections, vehicle A orients to the light andmoves in a straight course towards the source. Any deviation in course would slowthe motor on the side nearest the light and correct the vehicle’s path. Vehicle B,having crossed connections, moves in a fluttery pattern and always comes to restfacing away from the source. Thus vehicle B is easily snared by other lights.Given the light’s inhibitory influence on speed of approach and the outward-facinglight sensors, vehicle B often races off to distant lights.Note.Figure from Vehicles: Experiments in synthetic psychology. (p. 11), by V.Braitenberg, 1984, Cambridge, MA: MIT Press. Copyright 1984 by TheMassachusetts Institute of Technology. Reprinted with permission. Figure 2.The 15 ft x 15 ft arena was bordered by 12 in -wide shelves, set on theirsides and braced using clamps at the corners. Some of the lab activities requiredthat obstacles be placed in the arena. Obstacles were boxes ranging in size from ashoebox to a 14 ft square plastic crate. We covered each box in either tinfoil orblack construction paper, and placed a brick in each box to prevent the robots frommoving the obstacles. We set a few open boxes on their sides to create dark“caves”. Additional obstacles included an Astroturf doormat, sheets of sticky Understanding Behavior 18 rubberized shelf lining, and black posterboard ramps leading to inescapable pits(plastic tubs). Figure 3.SAM (Slightly Autonomous Machine) served as a class demonstrationrobot. SAM’s antennae activated touch sensors, allowing the robot to avoidobstacles. SAM also carried a forward-facing light sensor just above the antennae.The robot was powered by two motors, each connected to a separate rear wheel. Amotor and drive belt can be seen above the rear wheel at the right of the figure. Figure 4.Examples of robots built by students. Each robot carried an identicalprogram. Figure B1.Programs used for SAM and the students’ robots. Command “blocks”were dragged and dropped into place. The stack of commands on the left integratesthe individual behaviors programmed in the stacks to the right. “1”, “2”, and “3”refer to the computer’s input channels, and “A”, “B”, and “C” refer to outputchannels. There are three sensorimotor “rules” of behavior instantiated in theprogram. The first rule, under the “Dark” block, created a default movement in acounterclockwise, roughly circular path. The rule instructed the robot to turn onboth motors, then turn off the left motor every 1-2 seconds, and leave the left motoroff for a duration of 1-2 seconds. The second rule, under the “L-touch” and “R-touch” blocks, governed responses to objects. Upon activation of a left or righttouch sensor, the robot turned off the motor on the opposite side, resulting in a turnaway from objects. The third rule, under the “Light” block, created an attraction to Understanding Behavior 19 light. Whenever the light sensor returned a value over the threshold, the robotturned on both motors and emitted sounds. Understanding Behavior 20 Appendix A: Writing AssignmentsTopics for journal entries1. What mechanisms might be responsible for SAM’s behavior? Can you think ofpossible different mechanisms?2. What is the relationship between the goals, motives, and intentions we ascribe toorganisms and the mechanisms that make behavior?3. Can we just study minds? Or do we have to study “minds in bodies”?4. Is there an advantage to studying mechanisms in bodies situated in a physicalworld?5. How do multiple “minds” create social behavior? How can robot models helpus better understand the emergence of social behavior in living organisms?6. What are the implications of embodied cognition for the evolution of behavior?What gets acted on by selection? Rules in the head, physical structure, both,neither?EssayFrom a recent newspaper article (“In an Online Salon”, 1997):A crowd can empty a large football stadium in minutes, solving what is anintractable computational problem and exhibiting large-scale intelligence inthe absence of central direction. Why are decentralized processesubiquitous throughout nature and society – evolution is itself such a process– and why do people remain so distrustful of them?Discuss the relationship between this quotation, Braitenberg’s vehicles, and thedemonstrations in class. Understanding Behavior 21 Appendix B: Programming the RobotsWe programmed SAM and the students’ robots using the supplied RCXlanguage (version 1.0), a simple graphical environment in which we dragged anddropped commands to create “stacks” defining the relationships between inputs andoutputs (Figure B1). We found the programming easy to learn; an afternoon wasall that was required to create the programs and build a simple robot as a test bed.We used a PC for writing the programs (a Macintosh-compatible version of RCX isavailable through Dacta, the education division of Lego). The programs weredownloaded to the robots via an infrared port (included with the robot kit). SeeBaum (2000) for robot construction techniques and ideas, as well as additionalprograms in RCX and NQC, an alternative (and more powerful) C-likeprogramming environment.---------------------------------Insert Figure B1 about here---------------------------------