Abstract

We outline a possible theoretical framework for the quantitative modeling of networked embodied cognitive systems. We note that: (1) information self-structuring through sensory-motor coordination does not deterministically occur in ℝn vector space, a generic multivariable space, but in SE(3), the group structure of the possible motions of a body in space; (2) it happens in a stochastic open-ended environment. These observations may simplify, at the price of a certain abstraction, the modeling and the design of self-organization processes based on the maximization of some informational measures, such as mutual information. Furthermore, by providing closed form or computationally lighter algorithms, it may significantly reduce the computational burden of their implementation. We propose a modeling framework that aims to give new tools for the design of networks of new artificial self-organizing, embodied, and intelligent agents and for the reverse engineering of natural networks. At this point, it represents largely a theoretical conjecture, and must still to be experimentally verified whether this model will be useful in practice.

1 Introduction

In nature, cognitive adaptation is part of the overall adaptation process. On the one hand, a living being consuming less energy, all other conditions staying the same, is more likely to survive. On the other hand, in many ecological niches, and for the large majority of animals, the ones that move, the capability of processing more information in less time gives them a definite evolutionary advantage, as it allows them to move faster. Although there are other fitness criteria, depending on the species and the environments where they live, there is always pressure to improve energy efficiency and maximize information-processing capability. Moreover, shorter control programs have a higher probability of emerging from a random process than longer ones. As a consequence, we observe natural pressure to offload part of the information processing to the system dynamics. The exploitation of morphological computation schemes is a natural way to cope with these evolutionary pressures. These efficiency and effectiveness features—low energy consumption, high intelligence—are also needed by artificial intelligent systems: Many robots also need autonomy in the energetic sense, and simpler controllers are usually more reliable. Natural cognitive processes emerge from loosely coupled networks of embodied biological neurons; see, for a detailed discussion of this point, [5–7, 21]. It has also been shown that the successful adaptation of sensory-motor coordination, in a wide set of physical robotics settings, is usually characterized by a peak of the mutual (or multi-) information between the sensor and the actuators [22]. In [4, 8] it is argued that the recognition of the Lie group structure of the mobility space may help planning methods based on searching in the configuration space. It has recently been shown [4] that this allows us, given the statistical distribution of the joint variables to analyze, in certain cases, the controllability, observability, and stability of (some) kinematic chains from a Shannon information standpoint with compact closed form relations. Several researchers have shown the importance of information-driven self-organization (IDSO)—in particular Prokopenko, Der, and others [1, 15, 30], who used simulations of snakebots, humanoids, and grasping systems. These approaches seem very promising.

The combination of self-organization processes based on the maximization of suitable information metrics and the exploitation of the inherent structure of the motion of a macroscopic physical body might enable the design of effective and robust self-organized controllers and behaviors for sensory-motor coordination. It remains an open question whether this possibility is exploited or not in nature.

The main contribution of this article, which is of a rather hypothetical, conceptual, and theoretical nature, is to show that it is possible to design self-organizing emergent controllers of reduced computational cost by exploiting the robot morphology. This can be done by merging ideas coming from Prokopenko, Der, and others, with differential geometry and stochastic kinematics models developed, among others, by Chirijkian. The resulting method, proposed in this article, is also suitable for soft (deformable) robot systems and may be exploited in nature by animals.

Moreover, we review the basic ideas from IDSO, information theory, differential geometry, and stochastic kinematics, which might play an important role in the development of a quantitative theoretical framework for morphological computation.

The development of significant field experiments will be part of future work.

Despite the intuition and the valid arguments, shared by many researchers and scholars, that in nature some kind of embodied emergent self-organization process might be the general organizational principle of cognitive processes and of self and consciousness, the development of a quantitative framework for the modeling (and synthesis) of those processes has not been achieved so far; see [28, 29]. As an example, we may observe that while the Cornell Ranger [10] can walk for tens of kilometers on a single battery charge, there is no way to change its speed, which depends on the morphology of the system: to change speed, you must change the robot. A soft (variable impedance) legged system may in principle change speed, but how should the controller be shaped? How should you ground emergent embodied cognitive processes? How should you quantify the “performance,” or better, the fitness, of the different design options?

The modeling approach proposed here aims to provide an initial instantiation of a design methodology for the orchestration of emerging self-organizing controllers in soft robots and a possible model of similar processes happening in nature.

In what follows, we first review, in Section 2, some important results about the informational metrics of sensory-motor coordination. Then, in Section 3, we illustrate the concept of IDSO, and we discuss a number of examples from the literature. In Section 4, we show how the morphology of a physical body affects self-organization of information structures in physically embedded agents and how modeling these processes allows the design of more robust self-organizing controllers.

In Section 5, we analyze the conceptual and theoretical implications of this model, and we outline the open challenges and the future work.

We summarize, in a number of displayed passages, concepts and definitions that, while possibly obvious to readers with a background in information theory, differential geometry, or stochastic kinematics, might constitute a serious obstacle for the understanding of others.

The mathematical details are described in the appendices.

Definition 1 (Basic Information Metrics): The (Shannon) entropy offers a way to quantify the information carried by a stochastic variable with an associate probability distribution p(x):
formula
It can be proven that any metrics with the reasonable properties we want for an information metric must have a similar form. In particular, let us consider the following properties:
  • 1. 

    Continuity. The metric should be such that if you change the values of the probabilities by a very small amount, the metric should change by only a small amount.

  • 2. 
    Symmetry. The metric should not change if the results xi are reordered, that is,
    formula
  • 3. 
    Maximum. The metric should be maximal if all the outcomes are equally likely (uncertainty is highest when all possible events are equiprobable):
    formula
    For equiprobable events the entropy should increase with the number of outcomes:
    formula
  • 4. 

    Additivity. The amount of entropy should be independent of how the process is regarded as being divided into parts.

If a functional H has these properties, then it must have a form similar to that in Equation 1, that is,
formula
In a sense, Shannon entropy is the simplest way (K = 1) to associate a value to the information content of a stochastic variable. This makes it an interesting tool for the study of behaviors in an uncertain world. Another useful metric, in that it can be used to evaluate control systems in terms of how much they contribute to reduce the uncertainty on the state of the controlled variable [5, 6, 31], is the mutual information between two given variables
formula
where X and Y are two random variables.

If X and Y are statistically independent, the equation above gives I(X,Y) = 0 (X, the capital letter, represents the set of all the x values, and Y the set of all the y values).

A reference text on these and other topics related to information theory is Elements of Information Theory, by Cover and Thomas [11].

2 Informational Measures of Sensory-Motor Coordination

If we consider the analysis in [31], which shows how greater values of the mutual information between the sensors and the actuators characterize “good” controllers, the results in [22], summarized in Figures 1 and 2, are well understood. Lungarella and Sporns have applied some information metrics related to Shannon entropy and mutual information to quantitatively characterize the sensory-motor coordination of a set of physically different physical agents.

Figure 1. 

Evolution of some sensory-motor metrics for a number of physical settings, from [22] (courtesy of the authors). For all the three different experimental settings considered, there is a peak in transfer entropy when proper sensory-motor coordination is achieved.

Figure 1. 

Evolution of some sensory-motor metrics for a number of physical settings, from [22] (courtesy of the authors). For all the three different experimental settings considered, there is a peak in transfer entropy when proper sensory-motor coordination is achieved.

Figure 2. 

Experimental settings considered in [22] (courtesy of the authors). A1: a simple humanoid with an eye and an arm manipulating a ball. A2: a spider robot manipulating colored cubes. A3: a differential wheeled robot with a pan and tilt camera. B1, B2, B3: Schematic representations of the robots depicted in A1, A2, A3, respectively. C: The common learning and adaptation architecture implemented on the three different embodied agents (based on an artificial neural network connecting seasons and actuators). See [22] for details.

Figure 2. 

Experimental settings considered in [22] (courtesy of the authors). A1: a simple humanoid with an eye and an arm manipulating a ball. A2: a spider robot manipulating colored cubes. A3: a differential wheeled robot with a pan and tilt camera. B1, B2, B3: Schematic representations of the robots depicted in A1, A2, A3, respectively. C: The common learning and adaptation architecture implemented on the three different embodied agents (based on an artificial neural network connecting seasons and actuators). See [22] for details.

In Figure 2 the information flow (transfer entropy) between sensory inputs, a neural representation of saliency, and the actuator variables in the various physical settings, are represented. Transfer entropy, like Granger's causality, is a way to measure to what extent the future value of one time series is related to the current value of another; in some sense it quantifies how strongly a dynamic variable depends on another one. As mentioned earlier, the transfer entropy peaks when the system has learned an effective sensory-motor coordination schema (embodied in the weights in the neural networks). The environments used in [22] are intuitively very simple, and we would like to run experiments in more complex ones. What we need is a way to characterize the complexity of environments. But how to quantify how “cumbersome” an environment is? Lampe and Chatila [19] have proposed to measure the complexity of a simple environment by defining a metric based on the Shannon entropy. With reference to Figure 3, H is defined as the entropy related to the density of obstacles:
formula
where p (di) represents the ith density level in the occupancy grid, represented by the red square, with
formula
This metric seems to capture what we intuitively mean when we say that an environment is cluttered, as it gives higher entropy when the grid cells are occupied in a more random way.
Figure 3. 

How to calculate the entropy of a cluttered environment, from [19]. In the electronic version, the red square defines the atomic cell on which the density of obstacles is calculated. The density of obstacles is interpreted as the probability of finding an obstacle and is integrated to give the Shannon entropy of the obstacle distribution given by Equation 7.

Figure 3. 

How to calculate the entropy of a cluttered environment, from [19]. In the electronic version, the red square defines the atomic cell on which the density of obstacles is calculated. The density of obstacles is interpreted as the probability of finding an obstacle and is integrated to give the Shannon entropy of the obstacle distribution given by Equation 7.

We would, ideally, like to be able to understand a priori, for a given environment, the minimum complexity of an agent-embodied “brain” and system morphology dynamics to be able to perform a given set of tasks in that environment, relating this to its complexity. This might be connected to what Ashby called the “principle of requisite variety” [2]. The aim of the principles and the mathematical tools provide in this article is to go beyond a purely informational view and devise ways to include the morphology, the dynamics, and implicitly the materials in the model of the processes involved in the interaction of an embodied intelligent agent in its environment.

3 Information-Driven Self-Organization

The observations made above and experiments such as those summarized in the previous section suggest that the maximization of some metrics built on mutual information might be an important mechanism also at work in natural intelligent systems. The optimization of suitable informational measures might guide the emergence of cognitive processes and intelligent behavior in animals: This is what we refer to when we talk of information-driven self-organization. Information metrics, as we have seen in the previous section, are useful to study complex—in particular, multi-agent—information-processing systems. In what Prokopenko calls a strong form, IDSO could be regarded as one of the main drivers of natural evolution. The snakebot, by Tanev [30, 34], is an example of a system designed according to IDSO principles. It is interesting in that it shows that lifelike movements emerge by maximizing suitable information metrics. Another example is given by the hexapod walking model proposed by Cruse [12], where walking behavior emerges, without any central controller, through the interaction of the embodied system with the environment. The basic idea is that the parameters of a quite generic control system are tuned to maximize the meaningful interaction of the physical agent with the environment, and that the measure of this coupling is given by, for example, predictive information.

The snakebot is a simplified model of a rattlesnake, consisting of a series of loosely connected balls, see Figure 4. It can be shown that the amount of predictive information between groups of actuators (measured via a generalized excess entropy) grows as the modular robot starts to move across the terrain. The distributed actuators become more coupled when coordinated sidewinding locomotion becomes dominant. Note that if the states of the remote segments are synchronized, then some information has been indirectly transferred via stigmergy (due to the physical interactions among the segments, and with the terrain). As observed above, on the one hand the application of the predictive information maximization as a self-organizing method to generate behavior leads to outcomes that look realistic. On the other hand, the calculations involved are heavy, raising doubts on how it could happen in nature (and posing challenges to the utilization of these principles in the synthesis of artifacts). In [4, 8] it is argued that the incorporation of the Lie group structure, which characterizes themobility space of an embodied agent, may be helpful for planning methods based on searching in the configuration space. Stated differently: Including the body representation in the orchestrating controller simplifies the control. In this article, a method that on the basis of the previously quoted results may help the quantitative modeling of (natural and artificial) networked embodied systems is described. In Section 4 we show how this method can be applied to the evolution of sensory networks with emerging controllers and to the examples in Section 2; in Section 5 we discuss how this approach might be generalized to more challenging modeling contexts. In Section 3.1 we review, as an example of IDSO, a model of the evolution of sensory layouts in embodied agents, based on purely informational methods. We will reconsider this example when showing how to incorporate the morphology, through Lie groups, in the scheme.

Figure 4. 

Snakebot, by Tanev. (a) The snakebot, depicted as a series of loosely coupled balls, moving on a plane. (b) The snakebot moving between a number of obstacles. From [34] (courtesy of the authors).

Figure 4. 

Snakebot, by Tanev. (a) The snakebot, depicted as a series of loosely coupled balls, moving on a plane. (b) The snakebot moving between a number of obstacles. From [34] (courtesy of the authors).

3.1 Evolution of Sensory Layouts: Ashby's Proposal

Let us consider an example where an IDSO approach is used to evolve a morphology, in particular the morphological distribution of sensors in an embodied agent. It is called Ashby's proposal because the IDSO algorithms perform a pure optimization of informational metrics with no explicit consideration of the embodiment of the systems. The results are interesting in that they show that different morphologies of the sensors have different information-processing efficiencies—that is, morphology matters. In [25] a microbial genetic algorithm (GA; taken from Harvey [17, 24]) was used to evolve a sensory layout. The microbial GA is a bioinspired evolutionary algorithm that mimics the way microbes exchange DNA between different members of the population, horizontally, as opposed to vertically from generation to generation. It uses a steady state method rather than a generational method, meaning that instead of accumulating a complete new generation of offspring, and then eliminating all the members of the older generation and replacing it completely by the new, a single new offspring is generated at a time; then (in order to keep the population size constant) one member of the population dies, and it is replaced by the new one. The selection criterion can be implemented either by applying a fitness criterion to choose which parents will have an offspring, or by choosing which individual will die. The main benefits of the steady state over the generational criteria are that they are usually easier to implement and that they can be implemented in parallel. In summary, the microbial GA works as follows:

  • 1. 

    Pick two members of the population at random to be parents of the new offspring.

  • 2. 

    The less fit of the two parents is chosen as the one to die and be replaced.

  • 3. 

    Information can be transmitted horizontally within a generation.

Of course, any kind of optimizing strategy would fit here. An evolutionary programming strategy might be closer to what happens in nature, but it is possible that, for example, simulated annealing or some other method would work as well. The task for the evolved sensor layouts is to efficiently and effectively sense the environment which is captured by the fitness function ic, whose expression is given in Equation 9 below. Each individual's body is modeled by a 10 × 10 square with 10 sensors placed somewhere on it. The genome encodes a list of 10 positions within that square. At any generation, the individuals with the sensory layout maximizing a performance index based on informational measures are selected. The evolutionary algorithm has been tested on a number of different simplified environments; see Figure 5. The fitness function is given by a weighted performance index:
formula
This performance index balances redundancy through the mutual information between the sensors,
formula
and novelty through the Crutchfield's information metric
formula
where the conditional entropy H(Y|X) is expressed as
formula
Figure 5. 

The different environments used in [25] to evolve the sensor layout (courtesy of the authors). The environments must be sensed by an array of visual sensors, their geometric distribution on a flat 10 × 10 square is optimized in order to maximize their efficiency. The environment is modeled by a grid, a series of horizontal lines, a noisy colored background, and a picture representing real stones. In the electronic version, the red square is used as the atomic cell in which to calculate the metrics as in Figure 3.

Figure 5. 

The different environments used in [25] to evolve the sensor layout (courtesy of the authors). The environments must be sensed by an array of visual sensors, their geometric distribution on a flat 10 × 10 square is optimized in order to maximize their efficiency. The environment is modeled by a grid, a series of horizontal lines, a noisy colored background, and a picture representing real stones. In the electronic version, the red square is used as the atomic cell in which to calculate the metrics as in Figure 3.

It is worth noting that this is just one example of the many performance indices we may devise by weighting information metrics to balance exploitation and exploration. Figure 6 shows how the sensor layout approximates the distribution of single eyes in a composite insect eye: a hint in favor of strong IDSO. The evolved sensory layouts seem realistic. This suggests, as in the snakebot and similar examples, that we have good reasons to believe something similar is at work in nature. The main disadvantage, again, of this method is that it is computationally demanding. This example illustrates how “morphology matters,” as not all sensor layouts are equally fit, and eventually an insect-eye layout emerges. However, a limitation of the case shown is that it does not inherently take the morphology of the agent into account, because there is only a flat 10 × 10 square, and as a consequence, important aspects of the body morphology and kinematics are not considered.

Definition 2 (Different Metrics Related to Shannon Entropy): Because different researchers use different flavors of informational measures derived from the Shannon entropy, we summarize the ones we have quoted in this article:

  • • 
    Information flow or transfer entropy. Mutual information has some limitations when applied to time series analysis. As it is symmetric, it does not allow one to ascertain whether X influences Y or vice versa. In other words, it does not indicate the direction of the information flows. The transfer entropy [32], also known as information flow, circumvents this shortcoming. The transfer entropy is defined as
    formula
    The symbol h1 represents the entropy rate for the two systems, while h2 represents the entropy rate assuming that x is independent of y. We define the entropy rate as the amount of additional information needed to represent the next observation of one of the two systems.
  • • 

    Granger causality. The Granger causality, abbreviated sometimes as “G-causality,” is a form of causality based on statistical tests; see [16]. A stochastic variable X, or more specifically the time series of its sampled values, is said to Granger-cause the stochastic variable Y if the values in the time series of X influence the predictability of the future values of Y. It has been found by Barnett et al. [3] that transfer entropy and Granger causality are equivalent for Gaussian processes.

  • • 
    Excess entropy or predictive information. In general, predictive information represents the possibility to predict a future value of a time series when we know a series of past values. In a completely random time series, this quantity is zero. For Markov processes it is given by
    formula
    In other words, for Markov processes, the predictive information of a time series is equal to the mutual information between the current and the next measured values. This concept was actually proposed before by Crutchfield and Young in [13], with the name of excess entropy.

Figure 6. 

Metric projections of informational distances between visual sensors. Frames range from 1,800 in the top left picture to 3,000 in the bottom left picture. As time passes, the sensor layout is optimized. The V2 axis represents the informational distances in the two more statistically significant dimensions. As the agent moves from a simple environment, like the one represented by the horizontal lines, to a more complex one, like that represented by the rocks, the information distances between them assume the grid structure in the picture. This corresponds to different physical layouts in different environments. (From [25]; courtesy of the authors.)

Figure 6. 

Metric projections of informational distances between visual sensors. Frames range from 1,800 in the top left picture to 3,000 in the bottom left picture. As time passes, the sensor layout is optimized. The V2 axis represents the informational distances in the two more statistically significant dimensions. As the agent moves from a simple environment, like the one represented by the horizontal lines, to a more complex one, like that represented by the rocks, the information distances between them assume the grid structure in the picture. This corresponds to different physical layouts in different environments. (From [25]; courtesy of the authors.)

4 How to Deal with Embodiment

GAs, as population-based nongradient search methods, can be seen as search methods for high-dimensional spaces. The performance index maximized in [25], and the reinforcement learning method used in [30], are based on the brute force computation of informational metrics on the sensor and the actuator values. These metrics are in practice computed as binned summations of time series (i.e., you identify bins, a small number of consecutive time samples, and you compute the density of values representing the probabilities that are used in the calculations). The combination of an effective, but still computationally heavy, algorithm like the microbial GA (or in general any other optimization strategy) with a fitness function based on Shannon-like informational measures entails remarkable computational burdens. This has so far limited the applications to toy systems such as those discussed above and in [15, 25, 30]. Moreover, pure information-based metrics seem unsuitable for natural learning processes in natural intelligent agents, for the same reasons.

It is likely that the search process is simpler in nature, because part of the information processing is offloaded to the body dynamics. How can this be explained and represented in a formal way suitable to be exploited in an algorithm? The main idea put forward here is that the body shapes the computing in essentially two ways:

  • 1. 

    It reduces the available phase space to a well-defined subset of the possible movements in SE(3).

  • 2. 

    It exploits the symmetries in the possible motions.

Definition 3 (Symmetries, Lie Groups, and SE(3)): A symmetry in a system behavior (in particular in its motions) expresses the fact that the system does not change when subjected to certain changes; see, for example, Figure 7. Symmetries can be expressed mathematically with the concept of a group.

Figure 7. 

Examples of symmetry groups. The figure does not change if we rotate the hexagon by 60° or if we reflect the figure around the axis CF. Both of these operations belong to a symmetry group.

Figure 7. 

Examples of symmetry groups. The figure does not change if we rotate the hexagon by 60° or if we reflect the figure around the axis CF. Both of these operations belong to a symmetry group.

A group G is a set of objects that can be combined by a binary operation (called group multiplication or the composition rule, denoted by ○). The elements of the group are the objects that form the group (generally denoted by g). A generator is a (minimal) subset of elements that can be used to obtain (by means of group multiplication) all the elements of the group. More precisely, a group G is a set such that:

  • 1. 

    G is closed under multiplication (○)—if a,b are in G, then ab is also in G;

  • 2. 

    G contains an identity element I;

  • 3. 

    Each element g has an inverse g−1 (gg−1 = I) that is also an element;

  • 4. 

    The group composition rule (multiplication) is associative, that is, a ○ (bc) = (ab) ○ c (but is not necessarily commutative).

Figure 7 shows a discrete group in which the elements can be counted (i.e., it has an integer number of elements). More relevant to our problem are continuous groups. In a continuous group, the elements are generated by continuously varying a number (one or more) of parameters. Combinations of rotations and translations in space or on a plane can be represented by continuous groups (actually, they are instances of Lie groups). An example of a simple Lie group, shown in Figure 8, is formally called the group of all rotations in 2D space—the SO(2) group—and essentially represents a rotation of a (rigid) body on a plane:
formula
Figure 8. 

The group of continuous rotations in the plane. The group operation is given by the counterclockwise rotation of a point P1(x1, y1) around the origin of the coordinates by an angle θ to the new position P2(x2, y2). This operation can be represented by Equation 15.

Figure 8. 

The group of continuous rotations in the plane. The group operation is given by the counterclockwise rotation of a point P1(x1, y1) around the origin of the coordinates by an angle θ to the new position P2(x2, y2). This operation can be represented by Equation 15.

More formally, a Lie group is defined as a group whose elements can be parametrized by a finite number of parameters, that is, it is a continuous group that satisfies the two equivalent properties:

  • 1. 

    If g (ai) ○ g (bi) = g (ci), then ci is an analytic function of ai and bi.

  • 2. 

    The group manifold is differentiable.

For a given multi-rigid-body structure, such as an arm, a leg, or even a hand, the number of possible motions is limited to the composition of a comparatively small number of motions, represented by Lie (sub)groups. This fact greatly reduces the part of the system's phase space that has to be searched. As we will see in the next section and in  Appendices 1 and  2, for certain assumptions, it is also possible to approximate the informational metrics by means of closed form expressions. The ensemble of all the possible motions of a body in the usual tridimensional space (in general, compositions of rotations and translations) is called SE(3). Likewise, SE(2) is the ensemble of all the possible motions of a body on a plane (a bidimensional space).

4.1 Main Concepts

The main contribution of this article is the observation that it is possible to design information-driven self-organizing control processes by exploiting, through the Lie group formalism, the embodiment of the controllers in a physical macroscopic body. This is on the one hand more correct for the representation of the uncertainty connected with the embodied agent (see [20]), and on the other hand more computationally effective. It is more correct because if we apply the central limit theorem to a physical macroscopic body, we have a Gaussian in SE(3) for the g's, and if we marginalize in x,y,z, we have a distribution resembling a banana, not a Mexican hat (see Figure 9). This is more effective from the computational standpoint because, by exploiting the symmetries in the motion, we can in many cases dramatically reduce the burden of computing the informational metrics we want to optimize: The two situations in Figure 10 are exactly equivalent, a fact that is captured by the Lie group formalism. This methodology is made possible by merging mathematical results coming from two different lines of research, the IDSO line of research and the stochastic kinematics line of research. In summary:

  • 1. 

    We have banana, not Gaussian, distributions.

  • 2. 

    The bananas can be computed with (relatively) limited effort.

  • 3. 

    The optimizations can be performed by a population-based evolutionary programming scheme (among the many possible choices).

Figure 9. 

Statistical distribution on Lie groups. If we have a Gaussian multivariate distribution on a Lie group, the projection (marginalization) in the usual space is not a Gaussian, but a banana distribution, as we can see on the right side of the figure. In practice this means that, if we have a differential wheeled mobile robot moving toward a target, as in this picture, and we estimate its position at a given time, schematized by the line crossing the arrow, then we will estimate, if we assume a Gaussian distribution of position errors (pictured on the left), a nonzero probability to find it in a place where, if we consider the group structure of motion of a physical body (picture on the right), it has practically zero probability to be found.

Figure 9. 

Statistical distribution on Lie groups. If we have a Gaussian multivariate distribution on a Lie group, the projection (marginalization) in the usual space is not a Gaussian, but a banana distribution, as we can see on the right side of the figure. In practice this means that, if we have a differential wheeled mobile robot moving toward a target, as in this picture, and we estimate its position at a given time, schematized by the line crossing the arrow, then we will estimate, if we assume a Gaussian distribution of position errors (pictured on the left), a nonzero probability to find it in a place where, if we consider the group structure of motion of a physical body (picture on the right), it has practically zero probability to be found.

Figure 10. 

SE(2) symmetry. The two situations in this figure are identical. For example, when we calculate the predictive information for this mobile robot, we do not need to calculate it twice (and likewise for all the possible positions and orientations in SE(2)). A similar situation occurs in SE(3). These facts are captured by Lie groups.

Figure 10. 

SE(2) symmetry. The two situations in this figure are identical. For example, when we calculate the predictive information for this mobile robot, we do not need to calculate it twice (and likewise for all the possible positions and orientations in SE(2)). A similar situation occurs in SE(3). These facts are captured by Lie groups.

In the next sections, we first revisit the example introduced in the previous section in light of these concepts, and then we show how the transfer entropies in [22] can be computed more efficiently.

4.2 Revisiting the Evolution of Sensory Layout

In this section we describe an embodied version of the model of evolution of sensory layouts proposed in [25].

The possible motions of a physical body are structured in terms of Lie groups:

  • 1. 

    Articulated rigid multi-body systems: They are constrained to a finite group of rototranslations mathematically expressed by a finite number of Lie groups, subgroups of the general Lie group.

  • 2. 

    Deformable systems: We focus on infinitesimal motions. The possible motions of a material particle of a deformable continuous body are still constrained to rototranslations. Information metrics are computed on Lie groups with Lie algebra, instead of on flat ℝn spaces (the configuration space of a physical system can actually be regarded as a curved manifold embedded in ℝn).

The approach in [25], or similar works, can be modified when the algorithm is made aware of the body morphology. This can be achieved as follows. The evolutionary algorithm optimizes the performance index ic(S + C), a weighted sum of two metrics: one representing redundancy (through mutual information), and one representing the predictive power of the sensory-motor control system. Here S and C are the stochastic vector variables representing the state of the sensors and the controller (including the actuators); see [31]. There are three important differences in the approach proposed here, with respect to [25]:
  • 1. 

    We deal with emergent controllers embedded in a coevolving physical body structure (not only sensory layouts).

  • 2. 

    We analyze how the body shapes the controllers, by exploiting the powerful Lie group formalism and related concepts.

  • 3. 

    As we deal with controllers, we refer to a more suitable information metric: the predictive information.

This approach takes care of the limitations of the body dynamics, by considering the kinematic structure to which the system variables apply.

We may define a weighted performance index, weighting the predictive power of the overall system through the predictive information PI (a metric related to Shannon entropy) and the redundancy through mutual information between the different sensors:
formula
formula
Figure 11 shows how we can implement the suggested procedure by exploiting the (Lie) symmetries of the kinematic structure. The predictive information is given by Equation 50 in  Appendix 2 if we, for example, consider a kinematic serial chain made of a series of rigid bodies identified by a set of frames.
Figure 11. 

The approach considering embodiment. The performance index ic(S + C) used to guide the evolutionary algorithm is a function of the stochastic vector variables S representing the temporal series of the sensor values and C representing the temporal series of the vector values of the internal state of the controller and of actuation values; see [31] and [5] for details. Following the example in [25], we maximize a weighted sum of a metric representing redundancy in the system with a metric measuring the diversity. The second one is given by the predictive information, as we see sensing and actuation together. Both metrics are calculated considering the Lie group structure of motion by means of the formulas derived in the appendices and discussed in the main text.

Figure 11. 

The approach considering embodiment. The performance index ic(S + C) used to guide the evolutionary algorithm is a function of the stochastic vector variables S representing the temporal series of the sensor values and C representing the temporal series of the vector values of the internal state of the controller and of actuation values; see [31] and [5] for details. Following the example in [25], we maximize a weighted sum of a metric representing redundancy in the system with a metric measuring the diversity. The second one is given by the predictive information, as we see sensing and actuation together. Both metrics are calculated considering the Lie group structure of motion by means of the formulas derived in the appendices and discussed in the main text.

In general, the motion structure provides constraints so that the predictive information can be computed directly and in closed form. In addition, we can derive a closed form reinforcement learning rule maximizing it (with the assumption of tight Gaussian distributions), or (in more general cases without special assumptions) to simplify the computation. The concept of predictive information, also known as excess entropy, was introduced by Crutchfield and can be seen as a measure of complexity [13]. Der uses the time loop error as a measure of complexity [15] for time series. The details can be found in  Appendix 2.

Recognizing that the mobility space of a physical structure is actually a subspace of the Cartesian space ℝn has the potential, by dramatically reducing the computation cost, to make the computationally heavy IDSO methods applicable to nontrivial physical structures, by utilizing their greater limitations. The tradeoff is that this subspace is actually a curved manifold and that the operations on it are not commutative, yet a mature and powerful mathematical theory is available: Lie group theory.

4.3 Revisiting the Discussion in Section 2

The observations made above are true also when we want to analyze the informational tradeoffs between controllers and body dynamics. Let us consider again, for example, one of the physical instantiations of sensory-motor coordination briefly described in Section 2: A3 in Figure 2.

We have a differential wheel cart constrained to move on a plane, and we have a pan-and-tilt camera mounted on it. Thus, we have an SE(2) group symmetry and two rotational joints (corresponding to pan and tilt movements of the camera). We can then identify the motion group of the whole system as
formula
If we assume a Gaussian distribution on gGA3, the (differential) entropy will be given by Equation 38, and the predictive information by Equation 50.

Definition 4 (Useful Lie Group Operators and Properties of Rigid Body Motion): Exponential operators prove very useful in differential geometry, yet they are not widely known and used. Here, we review their basic definitions and properties. The Euclidean motion group SE(3) is the semidirect product of ℝ3 with the special orthogonal group SO(3).

We define an element g of SE(3) as , where aR3 and ASO(3). For any and , the group composition law is written as
formula
while the inverse of g is given by
formula
An alternative representation is given by 4 × 4 homogeneous matrices of the form
formula
In this case, the group composition law is given by matrix multiplication. For small translational (rotational) displacements from the identity along (about) the ith coordinate axis, the homogeneous transformation matrix is approximately given by
formula
where I4×4 is the identity matrix and
formula
Large motions can be obtained by exponentiating these matrices. It is useful to describe elements of SE(3) with the exponential parametrization:
formula
If we define the operator V such that
formula
as a consequence the total vector can be obtained as
formula

5 Discussion and Future Work

The examples proposed above can be more easily managed if we assume concentrated Gaussian distributions for the state variables and rigid multi-body dynamics. While the first assumption may not be particularly limiting in the context of controlled physical systems, the latter is not directly applicable to soft (deformable), natural, or artificial cognitive systems. In our perspective, to consider a deformable distributed controlled body as a rigid multi-body, one might be seen as a way of freezing degrees of freedom (DOF), which has been proposed as a viable control approach for soft robots. In other words: An emerging controller like those described here can be applied to embodied intelligent agents, which can be approximated by a rigid multi-body structure with concentrated uncertainties and elasticities, like the snakebot (Figure 4) or the ECCE robot (Figure 12); a fully distributed continuously deformable structure like that of an octopus will require more work, in particular from the mathematical standpoint.

Figure 12. 

The ECCE robot is a humanoid with a humanlike tendon-driven actuation of arms and hands and a deliberately imprecise body structure.

Figure 12. 

The ECCE robot is a humanoid with a humanlike tendon-driven actuation of arms and hands and a deliberately imprecise body structure.

It should be noted that, while in [25] the outcome of the evolutionary algorithm is a fit sensor layout, in systems like that represented in Figure 12, the outcome will be a fit configuration of the relative distances and orientations of, for example, eye and arm or hand, among themselves and from the sensors and the objects in the environment.

Are these methods or analogous ones exploited in nature? Is this the reason why the primate brain uses sometimes affine or quasi-affine geometries [23] for motion planning?

5.1 Interpreting the Freezing Mechanisms

It is reasonable to think that the DOF freezing mechanism might also be better understood and detailed from this perspective. For example, a wheeled mobile robot with a pan and tilt camera such as A3 in Figure 2, by freezing the pan and tilt motion of the camera, is reduced to SE(2)—a plane motion, the highly symmetrical mobility space depicted in Figure 9—and then it can more economically compute its moves.

Actually, the local properties of a deformable structure at a given time can be approximated infinitesimally by a rigid multi-body structure; from this perspective, the discussion above still holds, if we consider the physical approximating structure changing in time and if we see this approximating rigid body shape as part of the optimization process. In other terms, to locally decrease the stiffness of parts of a human arm during imprecise motion, for example in grasp preshape, might be seen as a way to transform a comparatively loose structure, with a high associated entropy, into a rigid multi-body structure (a “rag doll” in computer graphics jargon), with a considerably lower associated entropy and thus computational burden. On the contrary, during precise grasping, the additional DOFs provided by the deformability of the fingers make a proper grasp more likely.

5.2 A More General Formalism

These methods, when fully developed, will involve online optimization of the full sensing and actuation loop, based on embodied information-driven self-organization processes and exploiting the body's deformability to ease optimization. They may then have disruptive results, ranging from underactuated locomotion to softness of visual grasping systems. The generalization of this approach to fully deformable structures is part of future work and involves the application of a cleaner mathematical setting coming from differential geometry, and also from the theory of fiber bundles and connections [9, 18, 26, 33].

5.3 A Frame of Reference Problem?

Lie groups characterize the stochastic kinematics (and dynamics) of physical bodies, and we suggest that they should be incorporated in emerging controller schemes for sensory-motor coordination. We see in the simplification coming from the adoption of the Lie group modeling approach a palpable and measurable benefit of the incorporation of body morphology to ease computation. In other words, we regard it as a quite general and quantifiable example of morphological computation. As a consequence, as Lie groups are useful for a basic representation of the sensory-motor coupling of an embodied intelligent agent with the environment, an important aspect of the agent's learning will be the identification of the “structure of the space,” or rather the representation of what an agent can do with its individual body, its bodily affordances; see, for example, the work of O'Regan et al. [27], showing that the emergence of space awareness (abstract space representation) is a consequence of bodily affordances in an embodied agent.

6 Conclusions

In this article, we have shown a method to develop quantifiable information-driven, self-organizing sensory-motor coordination processes. This method may allow us to shape the emergence of control on the basis of the body morphology. This is possible at the price of greater abstractness and will require more work if the method is to be fully applicable to continuously deformable structures as exemplified by an octopus. It seems that something like what we propose here might be implemented to guide the freezing mechanism of a fully or partially deformable, natural, or artificial sensorial and actuation system.

A Lie group representation is a way to represent the body affordances (the body morphology and the morphological computation) to the “brain” of the artificial agent, without making a sharp distinction between the information processing and the dynamics.

Acknowledgments

The author thanks the reviewers and Prof. R. Pfeifer for their insightful and challenging remarks.

References

1. 
Ashby
,
W. R.
(
1960
).
Design for a brain: The origin of adaptive behavior.
New York
:
Wiley
.
2. 
Ay
,
N.
,
Bertschinger
,
N.
,
Der
,
R.
,
Güttler
,
F.
, &
Olbrich
,
E.
(
2008
).
Predictive information and explorative behavior of autonomous robots.
The European Physical Journal B—Condensed Matter and Complex Systems
,
63
(
3
),
329
339
.
3. 
Barnett
,
L.
,
Barrett
,
A. B.
, &
Seth
,
A. K.
(
2009
).
Granger causality and transfer entropy are equivalent for Gaussian variables.
Physical Review Letters
,
103
(
23
,
238701
.
4. 
Bonsignorio
,
F. P.
(
2007
).
Preliminary considerations for a quantitative theory of networked embodied intelligence.
In M. Lungarella et al. (Eds.)
,
50 Years of AI
(pp.
112
123
).
Berlin
:
Springer-Verlag
.
5. 
Bonsignorio
,
F. P.
(
2009
).
Steps to a cyber-physical model of networked embodied anticipatory behavior.
In G. Pezzulo et al. (Eds.)
,
ABiALS 2008
(pp.
77
94
).
Berlin
:
Springer-Verlag
.
6. 
Bonsignorio
,
F. P.
(
2010
).
On the stochastic stability and observability of controlled serial kinematic chains
In
Proceedings of the ASME 2010 10th Biennial Conference on Engineering Systems Design and Analysis, ESDA2010
.
7. 
Bonsignorio
,
F. P.
(
2012
).
The new science of physical cognitive systems: AI, robotics, neuroscience and cognitive sciences with a new name and the old philosophical problems?
In
SAPERE, studies in applied philosophy, epistemology and rational ethics.
Berlin
:
Springer-Verlag
.
8. 
Chirikjian
,
G. S.
(
2010
).
Information theory on Lie-groups and mobile robotics application.
In
Proceedings of the ICRA 2010
.
9. 
Chirikjian
,
G. S.
(
2011
).
Stochastic models, information theory, and Lie groups
,
Vol. 2
.
Berlin
:
Birkhauser
.
10. 
Cornell Ranger Web site
. .
11. 
Cover
,
T. M.
, &
Thomas
,
J. A.
(
2006
).
Elements of information theory
(2nd ed.).
New York
:
Wiley
.
12. 
Cruse
,
H.
(
1996
).
Neural networks as cybernetic systems.
Stuttgart
:
Thieme
.
13. 
Crutchfield
,
J. P.
, &
Young
,
K.
(
1989
).
Inferring statistical complexity.
Physical Review Letters
,
63
,
105
108
.
14. 
DelSole
,
T.
, &
Chang
,
P.
(
2003
).
Predictable component analysis, canonical correlation analysis, and autoregressive models.
Journal of Atmospherical Sciences
,
60
,
409
416
.
15. 
Der
,
R.
,
Martius
,
G.
, &
Hesse
,
F.
(
2006
).
Let it roll—Emerging sensorimotor coordination in a spherical robot.
In L. M. Rocha. (Ed.)
,
Artificial Life X
(pp.
192
198
).
Cambridge, MA
:
MIT Press
.
16. 
Granger
,
C. W. J.
(
1969
).
Investigating causal relations by econometric models and cross-spectral methods.
Econometrica
,
37
(
3
),
424
438
.
17. 
Harvey
,
I.
(
2001
).
Artificial evolution: A continuing saga.
In T. Gomi. (Ed.)
,
Evolutionary Robotics: From Intelligent Robots to Artificial Life, Proceedings of the 8th International Symposium on Evolutionary Robotics (ER2001)
(pp.
94
109
).
Berlin
:
Springer
.
18. 
Kolar
,
I.
,
Michor
,
P. W.
, &
Slovak
,
J.
(
1994
).
Natural operations in differential geometry.
Berlin
:
Springer-Verlag
.
19. 
Lampe
,
A.
, &
Chatila
,
R.
(
2006
).
Performance measure for the evaluation of mobile robot autonomy.
In
Proceedings of ICRA 2006
.
20. 
Long
,
A.
,
Wolfe
,
K.
,
Mashner
,
M.
, &
Chirikjian
,
G.
(
2012
).
The banana distribution is Gaussian: A localization study with exponential coordinates.
In
Proceedings of Robotics Science and Systems 2012
.
21. 
Lungarella
,
M.
,
Iida
,
F.
,
Bongard
,
J.
, &
Pfeifer
,
R.
(Eds.). (
2007
).
50 Years of AI, Festschrift.
Berlin
:
Springer-Verlag
.
22. 
Lungarella
,
M.
, &
Sporns
,
O.
(
2006
).
Mapping information flow in sensorimotor networks.
PLOS Computational Biology
,
2
(
10
),
1301
1312
.
23. 
Maoz
,
U.
,
Berthoz
,
A.
, &
Flash
,
T.
(
2009
).
Complex unconstrained three-dimensional hand movement and constant equi-affine speed.
Journal of Neurophysiology
,
101
(
2
),
1002
1015
.
24. 
McGregor
,
S.
, &
Harvey
,
I.
(
2005
).
Embracing plagiarism: Theoretical, biological and empirical justification for copy operators in genetic optimisation.
Genetic Programming and Evolvable Hardware
,
6
(
4
),
407
420
.
25. 
Olsson
,
L.
,
Nehaniv
,
C. L.
, &
Polani
,
D.
(
2004
).
Information trade-offs and the evolution of sensory layouts.
In J. Pollack, M. A. Bedau, P. Husbands, T. Ikegami & R. A. Watson. (Eds.)
,
Artificial Life IX.
Cambridge, MA
:
MIT Press
.
26. 
Olver
,
P. J.
(
2000
).
Applications of Lie groups to differential equations.
Berlin
:
Springer
.
27. 
Pfeifer
,
R.
, &
Scheier
,
C.
(
1999
).
Understanding intelligence.
Cambridge, MA
:
MIT Press
.
28. 
Pfeifer
,
R.
, &
Bongard
,
J.
(
2006
).
How the body shapes the way we think: A new view of intelligence.
Cambridge, MA
:
MIT Press
.
29. 
Philipona
,
D.
,
O'Regan
,
J. K.
,
Nadal
,
J. P.
, &
Coenen
,
O. J.-M. D.
(
2004
).
Perception of the structure of the physical world using unknown multimodal sensors and effectors.
In S. Thrun, L. Saul, & B. Schölkopf (Eds.)
,
Advances in neural information processing systems 16.
30. 
Prokopenko
,
M.
,
Gerasimov
,
V.
, &
Tanev
,
I.
(
2006
).
Evolving spatiotemporal coordination in a modular robotic system.
In S. Nolfi, G. Baldassarre, R. Calabretta, J. C. T. Hallam, D. Marocco, J. A. Meyer, O. Miglino, & D. Parisi. (Eds.)
,
From Animals to Animats 9: 9th International Conference on the Simulation of Adaptive Behavior (SAB 2006)
(pp.
558
569
).
Berlin
:
Springer
.
31. 
Touchette
,
H.
, &
Lloyd
,
S.
(
2003
).
Information-theoretic approach to the study of control systems.
Physica A
,
331
,
140
172
.
32. 
Schreiber
,
T.
(
2000
).
Measuring information transfer.
Physical Review Letters
,
85
,
461
464
.
Available at http://arxiv.org/pdf/nlin.CD/0001042.pdf (accessed July 2012)
.
33. 
Seiler
,
K. M.
,
Singh
,
S. P. N.
,
Sukkarieh
,
S.
, &
Durrant-Whyte
,
H.
(
2012
).
Using Lie group symmetries for fast corrective motion planning.
The International Journal of Robotics Research
,
31
(
2
),
151
166
.
34. 
Tanev
,
I.
,
Ray
,
T. S.
, &
Buller
,
A.
(
2005
).
Automated evolutionary design, robustness, and adaptation of sidewinding locomotion of a simulated snake-like robot.
IEEE Transactions on Robotics
,
21
(
4
),
632
645
.
35. 
Wang
,
Y.
, &
Chirikjian
,
G.
(
2006
).
Error propagation on the Euclidean group with applications to manipulator kinematics.
IEEE Transactions on Robotics
,
22
(
4
),
591
602
.

Appendix 1: Stochastic Kinematics of Rigid Bodies

We review here some known relations from the stochastic kinematics of rigid bodies. If we consider a vector function with , we can define μ by
formula
and Σ by
formula
We have the multivariable Gaussian distribution
formula
where
formula
In a similar way, we can define a function f(g) with gG, and two quantities, μ defined by
formula
and Σ by
formula
We have the multivariable Gaussian distribution on G:
formula
where
formula
This allows us to define a Gaussian distribution for the state variables; we have
formula
where , and g is defined as , with and ASO(3), and where can be obtained as x = (log g)V.
It can be shown [35] that if we define the matrix of covariances
formula
we have
formula
and
formula
The Shannon (differential) entropy associated with such a distribution is given by
formula
Let us now consider a kinematic serial chain made of a series of rigid bodies identified by a set of frames. It can be shown that given n shifted frames, with tight Gaussian distributions, we have a closed form expression for the quantities in Equation 39:
formula
with
formula
This allows us to compute the predictive information directly and in closed form and to derive a reinforcement learning rule maximizing it—at least in the previous hypotheses of tight (“concentrated” in technical language) Gaussian distributions.

Appendix 2: Expression of Predictive Information

It is now possible to derive a reinforcement learning rule with reference to known methods and relations from information theory; see [11, 14]. The expression for the predictive information in Markov hypotheses is given by
formula
For Markov processes, the predictive information of the sensor is equal to the mutual information between the current and the next measured values. Let us assume, then, that the state evolution can be modeled by assuming that the noise can be separated from the state vector, for each state vector component, as
formula
with the noise in ωi:
formula
We can then write, assuming a linear control function (a linear local approximation of the control system), the stochastic model:
formula
where W = N (0, Σ) with Σ given by Equation 36. If we assume the process is stationary, we have
formula
with
formula
We can express the distribution of the future sensor values in terms of the current values
formula
with
formula
It is then possible to express
formula
By applying the definition and exploiting the expression for multivariate normal distribution, we get
formula
with
formula
The control parameters are those for which we have
formula

Author notes

Robotics Lab, Department of System Engineering and Automation, University Carlos III of Madrid, Av. Universidad 30, 28911 Leganes (Madrid), Spain. E-mail: fabio.bonsignorio@uc3m.es