PI's: Sanjay Joshi (maejoshi@ucdavis.edu) , Department of Mechanical and Aeronautical Engineering, and Jeff Schank, (jcschank@ucdavis,edu) Department of Psychology
This material is based upon work supported by the National Science Foundation under Grant No. 0218927. NSF Disclaimer: Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the National Science Foundation (NSF).
Click on Topic
Overview -- Why Rat Pups? -- Methodology -- Rat Experiments -- Robots! -- Dynamic Simulation -- Results -- Personnel -- Papers
We are conducting a new biorobotics research program using Norway rat pups ages 7 to 10 days. This research is intended to inform both animal behavior and autonomous robotics. For animal behavior, the aim is to use robotics to overcome deficiencies in analytic and computational models used in the past. For robotics, studying infant mammals aims to enable the study of development of sensorimotor rules from the simple to the complex. Methodological goals of this project are to develop infrastructure and common scientific methods to conduct a parallel biology/engineering study of small mammals.
Norway rats (Rattus norvegicus) are a good example of an altricial mammal (i.e., born less developed and not well coordinated sensorimotor systems). Altricial mammals are relatively “simple” from a sensorimotor perspective since they are typically blind and deaf (e.g., rats, mice, dogs, cats) with limited motor and sensorimotor integration. Because they are extremely limited in their sensorimotor capabilities during the first two weeks postnatally, we can begin with relatively simple computational and robotic models of their behavior. However, they quickly develop new sensorimotor skills. This creates the opportunity to study group and individual behavior of an organism from a developmental perspective. In addition, rat pups are social animals from birth. As infants, they aggregate in huddles in which they can conserve energy and behaviorally thermoregulate. Aggregation is therefore an important group-behavioral function in Norway rats and many other species that produce multiple offspring. They are able to aggregate despite being blind and deaf at birth and for the first two weeks of life postnatally. This provides an opportunity to study multiple individual interaction, communication, and control.
How do we conduct our research investigations (Methodology)?
This is a truly interdisciplinary program utilizing tools, techniques, and ideas from robotics, engineering, dynamics, biology, psychology, and animal behavior. Our working hypothesis is that animal behavior in general, and specifically infant rat behavior, is dictated by four factors: (1) the mechanics of the animal including the body and its physical characteristics, (2) the internal state of the animal, (3) the environment of the animal, and (4) the rules of behavior implemented by the animal in an environment.
All the dependent variables in this hypothesis must be manipulated for a thorough study. However, all models (including laboratory experiments) have limitations. Individual-based probabilistic models used to discover simple rules in empirical data do not explicitly incorporate the physics of bodies or environments. Dynamic simulation models allow the explicit introduction of some known physical parameters for simulation, but are always based upon simplifications. Robots allow the explicit instantiation of physical parameters, but whether the instantiated parameters are relevant to the instantiated parameters of organisms is an empirical question. Thus, our approach is mandated by these concerns and unique in that animal, simulation, and robot studies occur in parallel and inform each other. Each line of investigation produces the same type of data about animal behavior, namely the position and orientation of the subject at any given time during an experiment.
A typical rat pup experiment consists of placing either a single pup or a group of pups in the middle of an arena, and observing behavior for a set period of time. Fig. 1 shows a single 7-day old rat pup in the arena. The arena is configured to study the effects of particular environmental cues (e.g. heat, inclines, etc.). It is very important that the test arena not be contaminated by uncontrolled environmental cues [e.g., light gradients, surface tilt, temperature gradients, vibrations, or odor gradients]. Even though pups are blind at birth and up to about 14 days of age, they can still detect light through their eyelids and with their developing eyes. Therefore, we have built a controlled environment that is far more precise in its ability to control factors affecting behavior than has ever been used in behavioral research with small infant mammals (shown in Fig. 2). The arena environment can be manipulated in a variety of ways for heat, light, and elevation. Above the arena, a camera records the motion of the rat pups.
A fundamental problem encountered in animal behavior research concerns how to collect and process sufficient amounts of precise data about animal motion. The equipment and methods must be non-invasive to avoid altering behavior. Thus, to extract precise data about the motion of our rat pups from video recordings, we developed an image analysis and data collection program. This program automatically stores a stack of static images from a video stream at a specified interval (e.g., every 5 seconds). The stack of images is then processed by researchers, who identify the head and tail location of each rat in the image. The result is position and orientation data of each rat pup in the arena over the entire experiment. An example of the kind of image created is shown in Fig. 3.
Figure 1: A single 7-day old rat pup moving about in the temperature controlled arena.
Figure 2: Experimental chamber used to study rat pup behavior.
Figure 3: A sample image used to record rat pup position and orientation. Researchers can track one or many rat pups by clicking head and tail locations over many still images.
Design and development of robotic rat pups has been an iterative process and is always continuing. The first two generations of robopups we designed and implemented are shown in Figs. 4 and 5. The mechanical design of our robots is motivated by the aim to emulate some of the relevant physical characteristics of rat pups. Characteristics we view as important include scaled rat shape, rat size, rat sensor locations, and rat paws for locomotion.
The shape of the robot is particularly important because it dictates how the robot interacts with corners and walls of the arena. Based on animal data collection, this is a defining feature of rat behavior. In general, rat pups are long with somewhat pointed heads culminating at the nose (Fig. 1). The rat pups use their nose to burrow into the crevasses and corners of the arena. This leads to a "jiggle" or "snooping" behavior. In rat pups, this behavior appears to sometimes lead to an escape from the corner, while sometimes a pup appears to be stuck in the corner for many minutes (or even the entire experiment). In order to emulate these shape features, we have created metal body skirts that are long and have tapered noses. Our behavioral observations have revealed that young pups (i.e., up to 10 days old), primarily use their two back legs for forward locomotion. Their front legs appear to help in orienting their head and body, but contribute little thrust for locomotion. As a result, rear driven locomotion (implemented on the robots with rear-driven wheels with differential drive on the chassis) creates forces and moments on the body that are similar to very young rat pups.
Sensors are mounted along the skirt. The controller and mechanical designs can accommodate several different types of sensors simultaneously. Our initial studies on very young rat pups are meant to explore their dominant modality - tactility. As a result, microswitch bump sensors are mounted at specific locations around the skirt. Since the majority of rat pup interactions occur around the rat head, we mount a number of sensors near the front of the robot. Our current generation of robopup can accommodate several closely packed sensors near the nose (Fig. 5). A few other sensors are mounted around the skirt to account for tactile contact between a rat's body and objects in its environment. The robots are controlled using an embedded microprocessor to implement behavioral rules connecting sensors to motors. Our current microcontroller is a Parallax Corporation Java-based STAMP microcontroller.
Figure 4: Original prototype of robotic rat ('the boat').
Figure 5: Roborat version 2.0.
Figure 6: Roborat in arena corner.
Figure 7: Eight roborats in test arena.
Simulation studies in our project serve at least three purposes: (1) They allow relatively rapid implementation of new rules to investigate robot and rat behavior, without needing to conduct robot experiments and collect data. (2) They allow studies to qualify and quantify the robustness of behavior to rule changes, environmental changes, and physical changes to robot models. (3) They allow for the possibility to run hundreds, thousands, or more simulated experiments (without human supervision) in order to study evolutionary methods for roborat controller design.
Study 1
Rattus Norvegicus is commonly referred to as a contact species (Barnett, 1963), since both activity levels as well as behavioral movements are shaped by contact with other objects as well as other agents (Schank & Alberts, 1997, 2000). This movement is largely thigmotaxic, meaning that they orient and move toward objects in their environment. Since at least 1934, thigmotaxis in the rat has been described as “reaction tendencies toward objects” (Patrick & Laughlin, 1934). Therefore, a natural starting point for a robot control system was a reactive architecture, responsive to contact alone, which implemented thigmotaxic behaviors reactively (Schank et al., 2004). Unlike real rat pups, robots implementing this reactive architecture produced very rigid and highly stereotyped behavior. For most runs, agents would quickly move to the wall, and then circumnavigate the arena for the remainder of the run. Therefore, the robotic rats always visited 4 corners, while actual rat pups sometimes visit 1, 2, 3, or 4 corners. Furthermore, rat pups spend a significant amount of time in the center area of the arena. The robotic rat pups did match some micro-level behaviors. Like rat pups, the robotic rats often followed walls (micro-level behavior). Also like rat pups, the robotic rats would often get stuck in corners (micro-level behavior). Nonetheless, it was clear that this architecture provided a poor model of rat pup behavior in an arena, and cast doubt on the purely reactive-thigmotaxic hypothesis for early rat pup behavior.
Study 2
In our second study, we showed the remarkable result that simply choosing robot movements at random, regardless of activation of sensors, produces quite intentional-looking emergent behavior patterns (Figure 8) that match actual rat pup patterns in both individuals and groups (May et al., 2006). We call these control rules non-sensory-dependent “random” rules. This result highlighted the strong affect of body and environment constraint on observed behavior. However, we could not conclude from these results that rat pups move about in an arena purely randomly because we had neither fully explored the dynamics of our rigid-body robots, nor investigated the space of possible sensorimotor rules.
Figure 8: Comparison of random robots and real rat pups. (This photo appeared on the cover of Complexity Vol. 12 No. 1 2006)
Study 3
In our most recent study, we used evolutionary algorithms (EA) to investigate the space of possible sensorimotor rules by artificially designing sensory-dependent deterministic robotic rat pup controllers using macro-level topological fitness metrics (Sullivan et al., 2007). The evolutionary algorithms allowed us to explore a wide range of possible deterministic control solutions. We showed that the best artificially designed controllers produced individual emergent behavior that captured the behavior of a real rat pup. Furthermore, in addition to matching behavior patterns at the macro-level, the evolved controllers produced behavior that also matched at the micro-level—even though these behaviors were not explicitly selected. Therefore, we have now found that both “random” rules, and (complicated) deterministic rules can both mimic observed rat pup behavior for an individual. (We are still in the process of evaluating the deterministic rules for groups.)
From a robotics perspective, our EA methodology also allowed us to observe the beginnings of self-organization of sensorimotor pairings that allow complex maneuvers to emerge by exploiting the body-environment complex, without the need for a central mediator (Sullivan et al., 2007). Since the current robot does not have the ability to bend its rigid body to extract itself from corners, a “reverse-backup” response for the nose sensor was needed. For example our reactive-thigmotaxic architecture (Study 1) created a backing mechanism for the nose sensor that involved two consecutive sets of wheel commands. If the nose sensor was activated, the robot would first back straight out, and then rotate either left or right by a certain angle before proceeding forward again. This pattern would continue for several cycles. In this way, the robot could eventually escape corners. In the meantime, this behavior looked like “corner-snooping” (a typical micro behavior observed in rat pups). In some cases, the EA evolved a particularly clever solution that used a set of two sequential sensor-actuator pairings. In the first pairing, the nose sensor would trigger a backward arc motion whose purpose was to engage a sensor on the robot back-corner. This back-sensor would then trigger a rotational motion that would escape the corner. In this way, a complex maneuver emerged which exploited physical features of both the robot-body and the environment, but did not require a central mediator. Interestingly, this type of sensori-motor pairing coordination would occur later in the evolution of controllers, after the EA had found the less efficient “corner-snooping” methods. This amounted to the beginnings of self-organization in our artificial system, and may have implications in robot design.
Summary of Findings to Date
In summary, our work (1) discounted a common belief in purely reactive cognition for rat-pup thigmotaxic behavior (Schank et al. 2004, May et al. 2006), (2) highlighted the importance of body-environment constraint on observed behavior (May et al. 2006), (3) promoted the further exploration of “randomness” in the brain-body-environment complex (Schank et al. 2004, May et al. 2006), (4) showed the value of evolutionary methods in the creation of feasible sensorimotor rules for emergent systems (Sullivan et al., 2007), and (5) began the exploration of self-organizing cognition for our artificial system (Sullivan et al., 2007). Our current work continues to explore these areas.
Current and Past Students:
Randall Bish (M.Sc., Mechanical and Aeronautical Engineering)
Mary Anne Canites (Undergraduate, Psychology)
Seth Egger (Undergraduate, Psychology)
Sobranie Frank (Undergraduate, Psychology)
Jan-Michael Galinato (Undergraduate, Psychology)
Nick Giannini (Undergraduate, Mechanical and Aeronautical Engineering)
Lisa Hargreaves (Undergraduate, Mechanical and Aeronautical Engineering)
Lauren Houston (Undergraduate, Psychology)
Marisano James (M.Sc., Computer Science)
Fred Lawton (Undergraduate, Electrical and Computer Engineering)
Kevin Leung (M.Sc., Electrical and Computer Engineering)
Ledenilla Manlogon (Undergraduate, Psychology)
Geoffrey Mante (Undergraduate, Psychology)
Chris May (Ph.D., Psychology)
Aisha Mitchell (Undergraduate, Psychology)
I. Scott (Undergraduate, Psychology)
Lindsey Spice (Undergraduate, Psychology)
Chris Sullivan (Undergraduate, Mechanical and Aeronautical Engineering)
RJ Taylor (Undergraduate, Psychology)
Jonathan Tran (M.Sc., Electrical and Computer Engineering)
Jason Wexler (Undergraduate, Mechanical and Aeronautical Engineering, UC Berkeley)
Joshi, S. and Schank (2003), J., Of Rats and Robots: A New Biorobotics Study of Norway Rat Pups, Proceedings of the 2nd International Workshop on the Mathematics and Algorithms of Social Insects, Atlanta, Georgia, December, pp. 57-63.
Joshi, S., Schank, J., Giannini, N., Hargreaves, L., and Bish, R. (2004), Development of Autonomous Robotics Technology for the Study of Rat Pups, Proceedings of the IEEE International Conference on Robotics and Automation, New Orleans, LA, April, 2860-2864.
Schank, J., May, C., Tran, J. and Joshi, S. (2004), A Biorobotic Investigation of Norway Rat Pups (Rattus Norvegicus) in an Arena, Adaptive Behavior, 12 (3/4), 161-174.
May, C., Schank, J., Joshi, S., Tran, J., Taylor, R., and Scott, I. (2006), Rat Pups and Random Robots Generate Self-Organized and Intentional Behavior, Complexity, 12 (1), 53-66.
Bish, R., Joshi, S., Schank, J., and Wexler, J. (2007), Mathematical Modeling and Computer Simulation of a Robotic Rat Pup, Mathematical and Computer Modeling, 45, 981-1000.
Sullivan, C., Joshi, S. and Schank, J. (2007), Evolutionary-Algorithm-Based Cognitive Design for an Individual Robotic Rat Pup, Submitted.