Reflexive and Cognitive Knowledge Representation in Autonomous Robots
A Project Funded by:
(Project Number: DUE-9650002)
The University of New Mexico PURSUE program
(Project Number: PP-71-99F)
and
Western Alliance To Expand Student Opportunities (WAESO)
Principal Investigator:
Project Description
I. Introduction
Pathfinder, the first of a series of NASA Discovery missions, demonstrated the feasibility of using low cost, self-contained robots for exploration of remote locations. The design goal for robots like the Pathfinder’s Sojourner was to create reliable systems that had enough computation power to navigate themselves within a dynamic environment while having as little complexity as possible.
In 1987, Rodney Brooks developed a concept for an insect-like robot that explored its environment reflexively using a subsumption architecture. A subsumption architecture is a collection of behavior modules arranged in a hierarchy where complex behavior emerges from combination of simple behaviors. One of the key features of this architecture is the absence of explicit knowledge representation. The robot cannot make a plan or remember it’s environment but rather interacts reflexively with its surroundings.
Along similar lines, Mark Tilden introduced BEAM robotics (Biology, Electronics, Aesthetics, Mechanics). BEAM robots, like Brook’s subsumption robots are reflexive. They navigate through the environment without an internal map. One of the overriding principles behind BEAM robots is that the machines must be self-sustaining and therefore must be quite simple. Most of the insect-like robots are solar powered and their behaviors are built into simple circuits without a CPU or memory.
In the purest form, the belief behind both architectures has been to develop robots using natural evolution and architectural simplification as guides. The robots are intended to first accomplish nature’s most basic animal task, survival in a dynamic environment, using as simple an architecture as possible. With some degree of success both approaches have met these goals. It now seems appropriate to move onto the next evolutionary step, to develop a robotic system that is both reflexive and reflective, that is, the system has the abilities to both reason about its environment and to successfully communicate its reasoning to other similar robots. A number of interesting questions emerge. First, what can and cannot a purely reflexive robot do? What is the next level above reflex? Is there a middle position – a mostly reflexive robot that is able to somehow act on the basis of experience, for example, as if it had a map of its environment? How would an internal reflexive (haptic) map of the environment be implemented? And finally, what is the role of inter-robot communication at the reflex level.
The goal of this project is to push reflexive robots one step up the evolutionary ladder by exploring what environmental knowledge can be represented and manipulated at the reflex level and what knowledge should best be represented at more complex, computational levels. The approach will be to integrate simple hard-coded BEAM robots and modular software-based subsumption robots.
To illustrate the direction of this exploration, consider two examples of reflex behavior, an infant learning to use his or her hands and a honeybee communicating information to the hive. When observing a newborn child, arm and hand movements appear to be mostly random, unguided behavior. Within this behavior though, certain rudimentary patterns of movement emerge, suggesting that the child already possesses certain fundamental motor knowledge and internal goals that direct the behavior. Within a very short period of time, the rudimentary motor skills become refined and are combined to form more complex goal-driven movement.
Now consider a honeybee. For general discussion it could be said that its behavior is primarily reflex driven, lacking reasoning and planning. Yet, the bee is capable of communicating information about food and directions to other bees in the hive. On some level it has stored a memory about where food is located and communicated that information to fellow members of its species who, themselves, store the information as a memory and use it to navigate back to the food. This brings up some interesting questions. What does the bee extract from its environment so that it can communicate? Does the bee need to make inferences about what it has extracted in order to either communicate or receive communication? If it does make inferences, what are they and how are they made? If the bee does not posses the ability to reason using abstract symbols, is it then perhaps using some sort of pre-symbolic processing?
From these two examples come the types of questions and concerns that will drive this research project. How much and what kinds of information can be represented at a reflexive level, where reflexive in this project means at the level of simple, non-computational circuitry? How can such information be processed to guide a robot’s behavior? At what point does such a reflexive approach, even if technically feasible, become inefficient and call for a more symbolic, even if neural net-like, computational approach? What are the appropriate levels, starting at the purely reflexive and moving through various stages of ever increasing computational and representational power, for mapping the environment and making plans?
Other related questions include: Should robots be like newborns and start with only rudimentary goals and behaviors in order to achieve maximum effectiveness in unknown hostile environments (i.e. to avoid being overspecialized and thus unable to adapt)? How newborn should the robots be, how much is nature and how much is nurture? How should new behaviors be learned given the initial set of goals and behaviors? How are learned behaviors passed on to other robots? And, interestingly, should some robots perhaps start off with special abilities and evolve into specialist?
II. Architectural Levels
The object of this research is to determine at what level should various goal-directed reasoning and behaviors be built into a robot. Consider a specific problem. A small walker robot is attracted to light and always attempts to move toward the brightest source. The robot is placed in a darkened room with an open door leading to a brightly-lit corridor. A brick is placed between the robot and the doorway. If the obstacle-avoidance behavior of the robot is to back up, turn left, take a few steps forward and resume goal-driven behavior, then the robot will most likely bump into the brick a number of times until it managed to move around the left side of the brick and walk out into the hallway.
A BEAM robot might be expected to repeat this behavior every time it was placed in this environment because of its inability to learn from its experience. A useful solution to this inefficient behavior would involve some way for the robot to recognized the situation and apply a solution gained from past experience. With the objective of keeping behavior as reflexive as possible, how much of this process could be handled at the BEAM level (reflexes hard-wired into the circuit)? Like the honeybee, at what level could a map or knowledge representation be used to make decisions?
III. Methods
With the objective of keeping all behavior as fundamental as possible, what kinds of representations are possible without the requirement of a processor and memory? How much can be represented at a pre-symbolic level in terms of mapping, planning, communicating and learning? To explore these questions a research team will be broken into two groups. Group one will be lead by an Engineering professor and will work on BEAM-type robots (that is to say, reflexive robots with behavior hard-wired into their electronic circuits) to develop solutions to the problem mentioned above. Group two, led by the Principal Investigator, a member of the Computer Science Department, will work with a subsumption architecture that incorporates some explicit representation of the environment and procedures to act on that representation.
To explore a subsumption architecture in the light-seeking problem discussed above, a Nomad Scout Robot will be used. The Scout is equipped with sixteen Polaroid Sonars, six touch-sensitive bumper switches, a light sensing photocell and two independent motor-driven wheels. The onboard processor is a Motorola MC68332 which is responsible for communicating, through a serial port, to a Linux laptop computer. The system, though designed to handle high-level symbol manipulation, will be modeled to imitate a subsumption-type architecture.
Traditionally the Brooksian subsumption architecture uses a network of Augmented Finite State Machines (AFSMs)* as the most basic units. Each of these modules performs an independent, low level task such as controlling a particular sensor or actuator. Using the Scout robot, the most basic foundation level will be composed of one AFSM for each sensor and two AFSMs controlling the actuators, one for each wheel. To activate one of the sonar AFSMs or the photocell AFSM, a threshold will be assigned (distance and brightness level). The touch sensors will be activated by simply breaking the switch. The actuators will be assigned a fixed baseline velocity when activated.
The second level of AFSMs will act as computational controls on reflex behaviors. Activation of a sensor or combination of sensors will result in the elicitation of a reflex behavior, generally the activation of one or both actuator AFSMs. While the intention of the project is to keep most behavior as basic as possible, other layers of AFSMs may need to be added to the behavior hierarchy in order to accomplish more complex combinations of behaviors. The subsumption research, beyond behavior activation and modification, will focus on a specialized layer of modules called the ‘Representation Layer’. Internal representation of the environment will be instantiated at a symbolic level and processed as such. This Representation Layer will be the memory for newly acquired (newly constructed) AFSMs, implemented as a set of software modules.
The team working on BEAM robots will be taking a slightly different tact. Mark Tilden, one of the leaders of the BEAM robot movement, argues that there is no well defined model for how to build reflexive, adaptive behaviors in simple electronic circuits. The Computer Science department presently has a Solarbotics ScoutWalker 2.1 whose behavior is determined by the orientation of the most luminous object in its environment. Using an attached Solarbotics Sunseeker head the robot pivots in the direction of light and attempts to move in that direction walking on its four legs. Obstacles are detected by a set of touch antennas mounted on the front.
It is possible to change this behavior within the present circuitry. The task of the BEAM team will be to evolve the basic design of a Solarbotics ScoutWalker 2.1. In the light seeking problem mentioned above, these robots initially will always run into the obstacle. The question facing the team is how to get the robot to adapt on its own to the obstacle. Novel solutions will often include physical adaptations to the robot. For instance, circuitry may be added that allows the robot to change its center of gravity or the amount of leg lift. Other physical adaptations may include an active feeler that sweeps in front of the robot actively seeking obstacles rather than passively wait to be touched.
Other circuit level modifications could include the addition of variable capacitors and resistors which can be adjusted by the robot’s own experience. For example, resistance and/or capacitor values could be modified so as change the movement of individual legs and thus vary the direction in which the robot walks. The goal would be to see if such modifications can act as a reflexive map of the environment, allowing the robot to more efficiently negotiate that environment.
One of the areas that will be investigated are ‘escape sets’ or ‘levels of panic’. Lacking the ability to reason about the nature of an adverse situation, the BEAM robot will initially use one behavior to escape. Failure to escape results in an escalation of ‘escape desire’. This could actually be an overall surge of actuator activety or could simply be a change in escape behavior, as in attempting to step over an obstacle instead of moving around it. The most productive set of escape behavior will be selected based on experiments run under different environmental conditions.
To conclude, both teams will attempt to move toward and eventually merge with each other. The subsumption team, while working to develop higher level decision making, will set its sight on creating maps of its environment at as low a symbolic level as possible. This will allow its work to merge with that of the BEAM team. The BEAM team, once it has determined how far it can go without any computational processing, will strive to create a robot with minimal computational power.
IV. Anticipated Outcome
The outcome of the first phase of this work will be to have a BEAM style robot augmented with decision-making abilities, implemented at the symbolic level. One could implement this immediately but the goal of this research is to determine the appropriate levels for the various representational and computational needs of such a robot. A second phase, not proposed as part of this project, will be to explore communications among a society of such robots and the resulting social interactions. This work could be of importance to the NASA unmanned space program in the development of inexpensive robots or vehicles capable of autonomous, complex behaviors.