5.8: Reorientation without Representation

The robotics work of Grey Walter has been accurately described as an inspiration to modern studies of autonomous systems (Reeve & Webb, 2003). Indeed, the kind of research conducted by Grey Walter seems remarkably similar to the “new wave” of behavior-based or biologically inspired robotics (Arkin, 1998; Breazeal, 2002; Sharkey, 1997; Webb & Consi, 2001).

In many respects, this represents an important renaissance of Grey Walter’s search for “mimicry of life” (Grey Walter, 1963, p. 114). Although the Tortoises were described in his very popular 1963 book The Living Brain, they essentially disappeared from the scientific picture for about a quarter of a century. Grey Walter was involved in a 1970 motorcycle accident that ended his career; after this accident, the whereabouts of most of the Tortoises was lost. One remained in the possession of his son after Grey Walter’s death in 1977; it was located in 1995 after an extensive search by Owen Holland. This discovery renewed interest in Grey Walter’s work (Hayward, 2001; Holland, 2003a, 2003b), and has re-established its important place in modern research.

The purpose of the current section is to briefly introduce one small segment of robotics research that has descended from Grey Walter’s pioneering work. In Chapter 3, we introduced the reorientation task that is frequently used to study how geometric and feature cues are used by an agent to navigate through its world. We also described a classical theory, the geometric module (Cheng, 1986; Gallistel, 1990), which has been used to explain some of the basic findings concerning this task. In Chapter 4, we noted that the reorientation task has also been approached from the perspective of connectionist cognitive science. A simple artificial neural network, the perceptron, has been offered as a viable alternative to classical theory (Dawson et al., 2010). In this section we briefly describe a third approach to the reorientation task, because embodied cognitive science has studied it in the context of behavior-based robotics.

Classical and connectionist cognitive science provide very different accounts of the co-operative and competitive interactions between geometric and featural cues when an agent attempts to relocate the target location in a reorientation arena. However, these different accounts are both representational. One of the themes pervading embodied cognitive science is a reaction against representational explanations of intelligent behavior (Shapiro, 2011). One field that has been a test bed for abandoning internal representations is known as new wave robotics (Sharkey, 1997).

New wave roboticists strive to replace representation with reaction (Brooks, 1999), to use sense-act cycles in the place of representational sense-think-act processing. This is because “embodied and situated systems can solve rather complicated tasks without requiring internal states or internal representations” (Nolfi & Floreano, 2000, p. 93). One skill that has been successfully demonstrated in new wave robotics is navigation in the context of the reorientation task (Lund & Miglino, 1998).

The Khepera robot (Bellmore & Nemhauser, 1968; Boogaarts, 2007) is a standard platform for the practice of new wave robotics. It has the appearance of a motorized hockey puck, uses two motor-driven wheels to move about, and has eight sensors distributed around its chassis that allow it to detect the proximity of obstacles. Roboticists have the goal of combining the proximity detector signals to control motor speed in order to produce desired dynamic behaviour. One approach to achieving this goal is to employ evolutionary robotics (Nolfi & Floreano, 2000). Evolutionary robotics involves using a genetic algorithm (Holland, 1992; Mitchell, 1996) to find a set of weights between each proximity detector and each motor.

In general, evolutionary robotics proceeds as follows (Nolfi & Floreano, 2000). First, a fitness function is defined, to evaluate the quality of robot performance. Evolution begins with an initial population of different control systems, such as different sets of sensor-to-motor weights. The fitness function is used to assess each of these control systems, and those that produce higher fitness values “survive.” Survivors are used to create the next generation of control systems via prescribed methods of “mutation.” The whole process of evaluate-survive-mutate is iterated; average fitness is expected to improve with each new generation. The evolutionary process ends when improvements in fitness stabilize. When evolution stops, the result is a control system that should be quite capable of performing the task that was evaluated by the fitness function.

Lund and Miglino (1998) used this procedure to evolve a control system that enabled Khepera robots to perform the reorientation task in a rectangular arena without feature cues. Their goal was to see whether a standard result—rotational error—could be produced in an agent that did not employ the geometric module, and indeed which did not represent arena properties at all. Lund and Miglino’s fitness function simply measured a robot’s closeness to the goal location. After 30 generations of evolution, they produced a system that would navigate a robot to the goal location from any of 8 different starting locations with a 41 percent success rate. Their robots also produced rotational error, for they incorrectly navigated to the corner 180° from the goal in another 41 percent of the test trials. These results were strikingly similar to those observed when rats perform reorientation in featureless rectangular arenas (e.g., Gallistel, 1990).

Importantly, the control system that was evolved by Lund and Miglino (1998) was simply a set of weighted connections between proximity detectors and motors, and not an encoding of arena shape.

The geometrical properties of the environment can be assimilated in the sensorymotor schema of the robot behavior without any explicit representation. In general, our work, in contrast with traditional cognitive models, shows how environmental knowledge can be reached without any form of direct representation. (Lund and Miglino, 1998, p. 198)

If arena shape is not explicitly represented, then how does the control system developed by Lund and Miglino (1998) produce reorientation task behaviour? When the robot is far enough from the arena walls that none of the sensors are detecting an obstacle, the controller weights are such that the robot moves in a gentle curve to the left. As a result, it never encounters a short wall when it leaves from any of its eight starting locations! When a long wall is (inevitably) encountered, the robot turns left and follows the wall until it stops in a corner. The result is that the robot will be at either the target location or its rotational equivalent.

The control system evolved by Lund and Miglino (1998) is restricted to rectangular arenas of a set size. If one of their robots is placed in an arena of even a slightly different size, its performance suffers (Nolfi, 2002). Nolfi used a much longer evolutionary process (500 generations), and also placed robots in different sized arenas, to successfully produce devices that would generate typical results not only in a featureless rectangular arena, but also in arenas of different dimensions. Again, these robots did so without representing arena shape or geometry.

Nolfi’s (2002) more general control system worked as follows. His robots would begin by moving forwards and avoiding walls, which would eventually lead them into a corner. When facing a corner, signals from the corner’s two walls caused the robot to first turn to orient itself at an angle of 45° from one of the corner’s walls. Then the robot would make an additional turn that was either clockwise or counterclockwise, depending upon whether the sensed wall was to the robot’s left or the right.

The final turn away from the corner necessarily pointed the robot in a direction that would cause it to follow a long wall, because sensing a wall at 45° is an indirect measurement of wall length:

If the robot finds a wall at about 45° on its left side and it previously left a corner, it means that the actual wall is one of the two longer walls. Conversely, if it encounters a wall at 45° on its right side, the actual wall is necessarily one of the two shorter walls. What is interesting is that the robot “measures” the relative length of the walls through action (i.e., by exploiting sensory–motor coordination) and it does not need any internal state to do so. (Nolfi, 2002, p. 141)

As a result, the robot sensed the long wall in a rectangular arena without representing wall length. It followed the long wall, which necessarily led the robot to either the goal corner or the corner that results in a rotational error, regardless of the actual dimensions of the rectangular arena.

Robots simpler than the Khepera can also perform the reorientation task, and they can at the same time generate some of its core results. The subsumption architecture has been used to design a simple LEGO robot, antiSLAM (Dawson, Dupuis, & Wilson, 2010), that demonstrates rotational error and illustrates how a new wave robot can combine geometric and featural cues, an ability not included in the evolved robots that have been discussed above.

The ability of autonomous robots to navigate is fundamental to their success. In contrast to the robots described in the preceding paragraphs, one of the major approaches to providing such navigation is called SLAM, which is an acronym for a representational approach named “simultaneous localization and mapping” (Jefferies & Yeap, 2008). Representationalists assumed that agents navigate their environment by sensing their current location and referencing it on some internal map. How is such navigation to proceed if an agent is placed in a novel environment for which no such map exists? SLAM is an attempt to answer this question. It proposes methods that enable an agent to build a new map of a novel environment and at the same time use this map to determine the agent’s current location.

The representational assumptions that underlie approaches such as SLAM have recently raised concerns in some researchers who study animal navigation (Alerstam, 2006). To what extent might a completely reactive, sense-act robot be capable of demonstrating interesting navigational behaviour? The purpose of antiSLAM (Dawson, Dupuis, & Wilson, 2010) was to explore this question in an incredibly simple platform—the robot’s name provides some sense of the motivation for its construction.

AntiSLAM is an example of a Braitenberg Vehicle 3 (Braitenberg, 1984), because it uses six different sensors, each of which contributes to the speed of two motors that propel and steer it. Two are ultrasonic sensors that are used as sonar to detect obstacles, two are rotation detectors that are used to determine when the robot has stopped moving, and two are light sensors that are used to attract the robot to locations of bright illumination. The sense-act reflexes of antiSLAM were not evolved but were instead created using the subsumption architecture.

The lowest level of processing in antiSLAM is “drive,” which essentially uses the outputs of the ultrasonic sensors to control motor speed. The closer to an obstacle a sensor gets, the slower is the speed of the one motor that the sensor helps to control. The next level is “escape.” When both rotation sensors are signaling that the robot is stationary (i.e., stopped by an obstacle detected by both sensors), the robot executes a turn to point itself in a different direction. The next level up is “wall following”: motor speed is manipulated in such a way that the robot has a strong bias to keep closer to a wall on the right than to a wall on the left. The highest level is “feature,” which uses two light sensors to contribute to motor speed in such a way that it approaches areas of brighter light.

AntiSLAM performs complex, lifelike exploratory behavior when placed in general environments. It follows walls, steers itself around obstacles, explores regions of brighter light, and turns around and escapes when it finds itself stopped in a corner or in front of a large obstacle.

When placed in a reorientation task arena, antiSLAM generates behaviors that give it the illusion of representing geometric and feature cues (Dawson, Dupuis, & Wilson, 2010). It follows walls in a rectangular arena, slowing to a halt when enters a corner. It then initiates a turning routine to exit the corner and continue exploring. Its light sensors permit it to reliably find a target location that is associated with particular geometric and local features. When local features are removed, it navigates the arena using geometric cues only, and it produces rotational errors. When local features are moved (i.e., an incorrect corner is illuminated), its choice of locations from a variety of starting points mimics the same combination of geometric and feature cues demonstrated in experiments with animals. In short, it produces some of the key features of the reorientation task—however, it does so without creating a cognitive map, and even without representing a goal. Furthermore, observations of antiSLAM’s reorientation task behavior indicated that a crucial behavioral measure, the path taken by an agent as it moves through the arena, is critical. Such paths are rarely reported in studies of reorientation.

The reorienting robots discussed above are fairly recent descendants of Grey Walter’s (1963) Tortoises, but their more ancient ancestors are the eighteenth-century life-mimicking, clockwork automata (Wood, 2002). These devices brought into sharp focus the philosophical issues concerning the comparison of man and machine that was central to Cartesian philosophy (Grenville, 2001; Wood, 2002). Religious tensions concerning the mechanistic nature of man, and the spiritual nature of clockwork automata, were soothed by dualism: automata and animals were machines. Men too were machines, but unlike automata, they also had souls. It was the appearance of clockwork automata that led to their popularity, as well as to their conflicts with the church. “Until the scientific era, what seemed most alive to people was what most looked like a living being. The vitality accorded to an object was a function primarily of its form” (Grey Walter, 1963, p. 115).

In contrast, Grey Walter’s Tortoises were not attempts to reproduce appearances, but were instead simulations of more general and more abstract abilities central to biological agents,

exploration, curiosity, free-will in the sense of unpredictability, goal-seeking, sel-fregulation, avoidance of dilemmas, foresight, memory, learning, forgetting, association of ideas, form recognition, and the elements of social accommodation. Such is life. (Grey Walter, 1963, p. 120)

By situating and embodying his machines, Grey Walter invented a new kind of scientific tool that produced behaviors that were creative and unpredictable, governed by nonlinear relationships between internal mechanisms and the surrounding, dynamic world.

Modern machines that mimic lifelike behavior still raise serious questions about what it is to be human. To Wood (2002, p. xxvii) all automata were presumptions “that life can be simulated by art or science or magic. And embodied in each invention is a riddle, a fundamental challenge to our perception of what makes us human.” The challenge is that if the lifelike behaviors of the Tortoises and their descendants are merely feedback loops between simple mechanisms and their environments, then might the same be true of human intelligence?

This challenge is reflected in some of roboticist Rodney Brooks’ remarks in Errol Morris’ 1997 documentary Fast, Cheap & Out of Control. Brooks begins by describing one of his early robots: “To an observer it appears that the robot has intentions and it has goals and it is following people and chasing prey. But it’s just the interaction of lots and lots of much simpler processes.” Brooks then considers extending this view to human cognition: “Maybe that’s all there is. Maybe a lot of what humans are doing could be explained this way.”

But as the segment in the documentary proceeds, Brooks, the pioneer of behavior-based robotics, is reluctant to believe that humans are similar types of devices:

When I think about it, I can almost see myself as being made up of thousands and thousands of little agents doing stuff almost independently. But at the same time I fall back into believing the things about humans that we all believe about humans and living life that way. Otherwise I analyze it too much; life becomes almost meaningless. (Morris, 1997)

Conflicts like those voiced by Brooks are brought to the forefront when embodied cognitive science ventures to study humanoid robots that are designed to exploit social environments and interactions (Breazeal, 2002; Turkle, 2011).