EXERCISES
3.5 A CRITIQUE OF THE NOUVELLE AI
FIGURE 3.24 DAMN Architecture is an example of bidding-mechanisms where the behaviour module ‘votes’ to decides which action to be pursued.
other behaviours, whereas in bidding mechanisms, behaviours are totally independent of each other.
Rosenblatt developed the Distributed Architecture for Mobile Navigation project (DAMN). In this architecture, a set of behaviours which cooperate to control a robot’s path by voting for various possible actions such as control of heading angle and speed, and an arbiter decides which is the action to be performed. The action with more votes is the one which is executed. However, the set of actions is predefined and it had a grid-based path planning.
Behaviour-based architectures can be classified depending on how the coordination between behaviours occurs; (1) Competitive: in these architectures, the system selects an action and sends it to the actuators, it is a winner-take-all mechanism. Subsumption architecture, action-selection and bidding mechanisms are examples of competitive coordination, and (2) Cooperative: in these architectures the system combines the actions coming from several behaviours to produce a new one that is sent to the actuators. Motor schema and Braitenberg vehicle-3c are example of cooperative coordination.
Unified Behaviour Framework (UBF) was designed to allow a robot to use competitive as well as cooperative policies using a single architecture, as shown in Figure 3.25. UBF allows for seamlessly switching between disparate architectural policies by varying the arbitration techniques. Often a low-level controller and its behaviour is developed for a specific task; thus any future redesign or extending the utility appreciably to other tasks is severely limited. Since UBF allows for seamless switching across distinct architectural policies during runtime, it encourages reuse of behaviours, with various different modes, viz. random activation of an arbitrary behaviour, priority based selection, cooperative and semi-cooperative arbitration etc. UBFs modular architecture helps to simplify design and development by code reuse.
Bidding mechanisms and ASM attempt to put in some deliberative content in reactive architectures to design higher level behaviours. This can be seen as early attempts which led towards hybrid architectures which is discussed later in the chapter.
115
FIGURE 3.25 Unified Behaviour Framework (UBF), provides dynamism to a behaviour-based approach by adding various modes of arbitration, which allows for the running of cooperative as well as competitive architectural policies on the same controller encouraging reuse of code. Two arbitrations using UBF, (a)Monte Carlo, the robot alternates randomly between wander and two modes of obstacle avoidance, (b)Highest Priorityarbitration is more like traditional subsumption where the behaviour with the highest priority which can either be preprogrammed or acquired through voting gets exeuted.
Traditional AI Reactive Approach
Designed in sensor fusion, models, goals, plans action selection, schemas, behaviour
Basis in search binding problem
Information flow
choose next action concurrent actions Emergent
aspects
behavioural response goals and plans which are made on the go
Robustness serial action, a module failure leads to system failure
parallel action leads to graceful degradation
Intelligence coded in by the manufacturer or the programmer
happens in tandem as an overlap of perception and action
3.5.1 Issues with the nouvelle AI
Subsumption architecture led to a near real-time implementation for mobile robots and later bipeds and humanoids, however suppressing and inhibiting layers of control meant losing information which can be critical in hard real time implementations as aerospace and military hardware in battlefields. Motor schema relies on blending and hence doesn’t have any issues of loss of information, though it inherits all the issues of potential field implementation, particularly the issue of local minima. Behaviour-based approaches are motivated from the natural world and do not work well to mirror human-like intelligence.
There have been suggestions that to develop higher-level intelligence akin to human beings, a representation free approach will not prove to be sufficient.
3.5.1.1 Implementation issues in subsumption architecture
1. Loss of information and disruption of lower-level functionality:As lower-level functionalities are suppressed or inhibited, the information contained in those layers is lost. Hence a robot, when executing a higher-level behaviour, may fail to execute a simpler behaviour at the lower levels.
2. How to identify a higher level behaviour: Determining a given behaviour as higher level or lower level is a matter of arbitration [143] and there is no real methodology to ascertain this. The general rule of suppressing lower levels in favour of higher levels is not the hall mark of a good design.
3. Layers are not really independent of each other:Subsumption assumes that each layer of control is independent of another, but this is not always true. The internal
117
FIGURE 3.26 Lack of a compromise solution. Subsumption fails to identify a compromise solution, adapted from Rosenblatt and Payton [282].
functionalities of the higher levels often interfere with the lower levels, and often times two or more layers use the same hardware making independence of those layers really questionable. Thus small changes to the lower levels are difficult to implement without redesigning most of the other layers. In particular, higher layers can suppress, inhibit and also read the signals of lower layers. Adding higher level behaviours without altering lower level functionality is not possible. The physical connections between layers are hard-wired, so they cannot be changed during execution, making the system less dynamic to changes in the environment. As an example, consider a motor which is a common hardware for 2 layers of wandering and motion to goal. When wandering is suppressed by motion to goal, the motor is already at a non-zero velocity while motion to goal would have been designed considering the motor starts from zero. It may happen that at a higher velocity as attained during wandering, the motor may be difficult to align towards the goal point or, even worse, miss it all together.
4. Levels must have a priority: Each behaviour is associated with a degree of priority, thus behaviours with equal priority cannot be represented with the subsumption architecture.
5. Lack of access to the internal state of each behaviour:With increasing layers of control it is often difficult to acertain the information internal to a behaviour.
This shortcoming poses as a design issue and the coordination of a large set of behavioural skills by adding further layers of control to achieve a desired coherent complex behaviour is a difficult and error-prone job, lack of modularity is alarming in hard real time systems such as aircraft, elevators and medical applications. Critics have claimed that such designing of behaviours is apparently more of an art than a science.
could have been used as a consistency for both the behaviours, however subsumption doesn’t merit such compromise solutions. To overcome this lacuna, Rosenblatt and Payton [282] suggested a connectionist architecture which uses a very fine grained layered architecture. Instead of designing behaviours as finite state machines with internal states and instance variable, behaviours are comprised of atomic functional elements that have no inaccessible internal state. These simple decision making units and their interconnections collectively define a behaviour. Since the behaviours can interact with each other, thus can select a best compromise solution. In continuation of this work, Rosenblatt also develop a notion of ‘internalized plans’, instead of mere program for action, plans are considered as sources of information and advice to agents that are already accomplished in dealing with a dynamic environment. Therefore, plans are used selectively, and serve to enhance system performance or as fail safes in dire scenarios. In implementation, the notion of ‘internalized plans’ is realised using a global navigation using potential-field based gradient descent method.
3.5.1.2 Issues with motor schema
Instead of suppression or inhibition motor schema uses blending of behaviours akin to vector addition, so there is no loss of information. However, three issues of this approach are:
1. Local minima problem: A built-in shortcoming of motor schema is the issue of local minima which plagues all potential field implementations3.
2. Lack of practical solution by blending: A problem that cooperative mechanisms such as motor schema face is that the solution via blending behaviours may not really solve the problem. Consider a robot which has a gaping hole straight ahead, and two different behaviours attempt to avoid it, one trying to avoid it through the right and the other one trying to avoid it through the left. The sum of the 2 behaviours would be a vector pointing straight ahead to the gaping hole, the emergent behaviour fails to solve this problem. Researchers have tried to overcome this problem using dynamic methods for behaviour selection and bidding strategies where the the option with the highest level of activation or the most popular action is executed.
3. Null summation: Another implementation issue is when two or more schemas of the same magnitude working in opposite directions to each other will nullify the net vector sum. Thus though there is stimuli but it does not evoke a motor response.
2The next chapter will include further details on ND navigation.
3The next chapter details the various implementation of potential field method for navigation and the methods to overcome the local minima problem will be discussed there.
119
FIGURE 3.27 ‘Today the earwig, tomorrow man?’ David Kirsh argues in favour of representations and maintains that recreating insect-like behaviour in robots is not sufficient to elicit higher levels of intelligence. Kirsh associates representations with perception, learning and control, thus making it imperative in development of human like intelligence. Brooks replied to this criticism by suggesting that the behaviour-based paradigm needs to incorporate cognitive approaches to develop humanoid robots with human-like faculties, Brooks named this discipline
‘cognobotics’.
3.5.2 Extending reactive approach to higher functions
Brooks’ hypothesis of ‘reason and representation free’ AI works very well for situationally determined activities such as walking, running, avoiding obstacles, climbing stairs, etc.
However it is silent regarding cerebral activities such as playing chess, solving jigsaw, tinkering with the Rubik’s cube, etc. For example Dennett’s robots; R1, R1D1 and R1D2 without any representative knowledge, viz. a visual knowledge of the battery and/or the bomb looks like, will all fail the task and [once again] will be bombed to destruction, this time on a purely behaviour-based approach.
Kirsh [180] in his opposition to reactive approaches, insisted that robotics can never be modelled without representations, such as development of ‘insect like’ behaviour such as wandering, avoiding obstacles and wall following will never be sufficient to progress to higher levels of intelligence, as shown in Figure 3.27. Kirsh agrees that 97% of human behaviour is nonrepresentational, however in the remaining 3%, human beings engage in cerebral activities, the typical non-trivial aspects of human intelligence which lack in reflexes and animal intelligence. Kirsh argues that there is a limit to which a creature can go without
‘concepts’4and for higher level intelligence there is a need to have conceptual representations in action, perception, learning and control [112].
Kirsh’s criticism however doesn’t suggest any other method to achieve artificial human-like intelligence. In similar context, Wilson [357] points to the lack of a model for complete understanding of intelligence — the lack of a strong relation between methods of natural science and nature.
The bottom-up approach suggested by the behaviour-based philosophy is often insufficient and fails to reach human-like intelligence. ANIMAT & ALIFE research attempts
4Kirsh prefers to use the term concept-free rather than representation-free. He suggests that mobile robots control truly concerns the conceptualisation in intelligent activity, rather than representation as modelled in reactive approaches.
FIGURE 3.28 Two modes of cognition. Higher life forms as human beings at least have two modes of cognition, reactive and deliberative which work in tandem.
to develop human-like intelligence by simulating and understanding complete animal like systems at a simple level. The shortcoming is most glaring in more sophisticated behaviours, and since cognition is not limited to reflexes in higher beings. Neuroscience also suggests two
‘modes’ of human behaviours as shown in Figure 3.28. Similar ideas have been suggested by Steve Pinker for two levels of consciousness in higher life forms such as human beings which will be discussed inChapter 9.
The Philosopher & The Robot
Agent-based robotics is seen as the pinnacle of cutting edge technology, while in contrast philosophy is perceived as the domain of the proverbial armchair, greying professors and old books. However, the two disciplines are very closely related. As discussed in the previous chapter, the problem of relevance and the contention of using internal representations is a philosophical debate rather than a technological limitation.
Philosophical concerns nearly always dovetail with agent-based robotics, and this will be seen in later chapters inquiring into ethical robots, artificial consciousness and super intelligence.
In the early 1960s, agent-based robotics was conceived more as an extension of the computer and therefore intelligence was approached as the Cartesian model of the
‘thinking thing’, and information processing was central to this model. The robot had to sense the local environment, make maps and find out the most suitable paths to the goal point, and then move these paths. This process of sense-plan-act was cyclically incremented to work in a continiuous manner. The onboard computer was the ‘brain’
of the robot and other sensors were input units. By processing the data from the sensors, the robot could draw in on representations to form Kantian concepts of space and time [150], and therefore have meaningful interaction with the environment. The
121
outcome was a very slow-moving robot with a bulky onboard computer which was not at all suitable for a dynamically changing environment with moving obstacles and changing terrain. The issue of this non-performance was attributed as the lack of fast processing units, and it was believed that the problem would dissolve with better technology and lead to the desired results.
In the early 1970s, Dreyfus pointed out the limitations of traditional AI by contesting the shortcomings of the computational paradigm in his seminal bookWhat computers cannot do, and demonstrated the philosophical pitfalls in this approach. His argument was based on the explosion in the number of symbols and rules to overcome the frame problem. Dreyfus suggested the use of the philosophy of Heidegger and Merleau-Ponty to overcome the lacuna of disembodied, acontextual and symbolic computation which was being pursued by traditional AI. He suggested a paradigm shift towards what he coined as ‘Heideggerian AI’, which abandoned the rational approach of representation and planning and was designed on the agent’s dynamic interaction with its local world and focussed on immediacy rather than long-term planning, and therefore intelligence was related to sensorimotor action rather than slow and incremental designs of plans.
This consistent coping with the environment has been namedBeing-in-the-Worldin the works of Heidegger. In the behaviour-based paradigm, the concepts of situatedness and embodiment are directly drawn from Heidegger and Merleau-Ponty, respectively.
Heideggerian AI is able to avoid the inherent problems of traditional AI as it goes around both the frame problem — by being situated — and also the symbol grounding problem
— by being embodied. Heideggerian AI is the philosophical basis for connectionist approaches to cognition and behaviourist approaches to artificial life. Other than robotics, Heideggerian AI can also be seen in Agre’s development of indexical functional entities to design the dynamics of PENGI and Freeman’s neurodynamic model of the brain. In the 1990s, Varela and co-researchers [320, 333] developed the enactive paradigm by extending Merleau-Ponty’s phenomenology of the body to embodied AI.
The approach highlights that the agent’s body is not only a lived-in and experiential structure, but it is also the backdrop of all cognitive processes.
In the attempt to develop human-like intelligence, the behaviour-based paradigm is an exposition into the minimal conditions needed for a humanly experience [92].
However, this is not sufficient to lead to human-like intelligence, and neither ethical agency nor conscious behaviour can be extrapolated from it. This will be discussed in Chapters 9 and 10. As of now, most state-of-the-art robotic systems employ a twofold hybrid system: low-level behaviours are attributed with a reactive module and higher-level behaviours are controlled with a deliberative module.
In the late 1990s Brooks extended behaviour based approach to humanoids with the making of his robot Cog, and coined it as a ‘second shift’ [55] in behaviour based paradigms, a leaning towards cognitive robotics. He believed that bodily form and morphology of the human like robot, development of artificial motivation — which will naturally prioritise events, coherent action of various smaller subsystems and adaptation to motor control
& world dynamics will play important roles in the ‘second shift’. Development of such
‘cognobots’ will have to be channeled through an integrated approach where the cognitive development, sensor engineering and integration and morphological design work in parallel and the various subteams are aware of the workings of each other. It is imperative that study of the human brain was meant to greatly influence cognobotics. Cog was planned as a long-term project and progress has slowed down with passing years, but it still has served
philosophy inFigure 3.29. For example, I have discussed the design of a foraging robot using subsumption architecture, which is ispired from similar activities in the animal kingdom.
For a single deer foraging for food, it is guided by its reactive sense of olfaction however it is also led by visual stimulus as it knows what the food looks like (viz. leaves, herbs, shrubs etc.), knowledge it has acquired from experience. Cognition is not left to a solitary senor-actuator pair but it rather happens in tandem across a number of sensory organs.
Similar arguments can be made in favour of birds, rats etc. Animals, birds and even some insects acquire representational knowledge through experience which is useful in tasks as navigation, foraging, predator-prey scenarios etc. For human beings, psychologists have ascertained that human behaviour works across two distinct modes, willed and automatic.
Reactive architecture are often meant for a specific set of tasks. A controller designed on the reactive paradigm cannot easily be made to work with another set of tasks without modifying nearly the entire controller. In contrast, deliberative agents work well in static environments and fare poorly in dynamic environments. Therefore, a robot made to do complex tasks in a dynamic world cannot accomplish them in a purely reactive manner nor in a deliberative plan-based manner, hence there is a need to look into methods to combine the goodness of both of these. An apparently minimalistic solution is to equip behaviour based system with maps. Sensor based knowledge of the immediate neighbourhood is often realised with grid representations, while a priori maps generally have more global data.
The third way to incorporate knowledge is through perception. Such systems are usually equipped with state-of-the-art vision systems.
Reactive approach with maps strictly pertains to navigation, so development of systems which have identifiable architectural components of both deliberative and reactive paradigms have found more acceptance in the AI community. The two paradigms can be combined in various ways. The simplest designs attempt to redesign the sense-plan-act approach to adhere to a reactive paradigm. In it the robot would generate a plan to accomplish a job and execute it reactively in a sense-act execution, after which the plan module will generate further plans, while multilayered approaches have been popular with researchers, as shown inFigure 3.30.
Hybrid architectures have a striking similarity with a managerial corporate heirarchy with high level planning and policy decisions done at the higher up ‘manager’ levels, which then pass them off for execution to the lower levels, while the lowest level, the reactive level, is a corollory to the ‘worker’, which executes the job and is more well acquainted with the immediacy of the job at hand.
The works of Firby, Gat, Connell, Bonasso etc. led to the development of the three layer architecture [116] and its variants — 3T architecture, ATLANTIS and SSS. Behaviour based route to controlling robots considers direct mapping of sensors onto actuators in contrast the 3 level architecture work on two types of algorithms, (a) algorithms for governing routine