1. Introduction

Robots are being extensively used for the purpose of discovering and testing hypotheses about biological sensorimotor behaviors. This use of robots is the distinguishing feature of what is now being called “biorobotics” (Webb Reference Webb2001, Reference Webb2006; Webb and Consi Reference Webb and Consi2001). Significant examples of biorobotic inquiries include the investigation of phonotaxis in crickets (Webb Reference Webb2002), chemotaxis in lobsters (Grasso et al. Reference Grasso2000), ant homing capacities (Lambrinos et al. Reference Lambrinos2000), and hippocampal navigation in rats (Burgess et al. Reference Burgess1997). Biorobotic approaches are adopted to study human sensorimotor systems too: arm posture maintenance is investigated by implementing the hypothesized neural control mechanism onto an artificial arm equipped with a robotic muscle spindle (Chou and Hannaford Reference Chou and Hannaford1997; Jaax and Hannaford Reference Jaax and Hannaford2002); primate optomotor systems are investigated by means of VLSI implementations (Horiuchi and Koch Reference Horiuchi and Koch1996); the role of basal ganglia in action selection (Prescott et al. Reference Prescott, Nicholson and Faulls2002), characteristic patterns of human arm trajectories (Schaal and Sternad Reference Schaal and Sternad2001), cerebellar control (Eskiizmirliler et al. Reference Eskiizmirliler2002), and visuo-motor coordination (Dario et al. Reference Dario2002) are being explored on similar grounds. All of these inquiries involve experimental comparisons between biological system and robotic system behaviors.

Biorobotics shares with artificial intelligence and so-called “biologically-inspired” robotics the goal of achieving adaptive behaviors in machines. Biologically inspired robotics adopts solutions to robotic control problems suggested by models of adaptive biological behaviors only insofar as these solutions enable one to improve the adaptivity of service robots (Beer et al. Reference Beer1997). Biorobotics has to meet more stringent methodological demands, insofar as robots serve there the special purpose of testing or discovering theoretical models of biological sensorimotor behaviors. The potential contribution of this research strategy to the scientific understanding of biological systems was emphasized as early as the beginning of last century, when American biologist Jacques Loeb claimed that newly constructed phototropic machines provided empirical support for his model of phototropic behavior in moths (Loeb Reference Loeb1918; Cordeschi Reference Cordeschi2002). More widely acknowledged ancestors of biorobotics include pioneers of Cybernetics Arturo Rosenblueth, Norbert Wiener, and John Bigelow. In articles published in this journal (Rosenblueth, Wiener, and Bigelow Reference Rosenblueth, Wiener and Bigelow1943; Rosenblueth and Wiener Reference Rosenblueth and Wiener1945), these scientists conjectured that a variety of goal-directed behaviors observed in biological systems are governed by negative feedback control mechanisms, and suggested on these grounds the opportunity of using feedback controlled machines as material models of biological systems: these material models would enable one to generate empirical predictions that are difficult to work out on the basis of theoretical models only, and to make interventions that are difficult or unethical to perform on biological systems.Footnote ¹

Distinctive methodological problems arise in the process of designing and performing experiments with machine models. The methodology of machine experiments is examined here in connection with contemporary biorobotic investigations of biological sensorimotor mechanisms. In these investigations one has to address the problem of constructing or licensing the use of robots for experimental purposes. More specifically, one has to map biological mechanism descriptions into robotic mechanism descriptions, through the intermediary of some “mechanism schema” (Machamer, Darden, and Craver Reference Machamer, Darden and Craver2000) specifying abstract role functions that are filled in by means of component parts and processes of biological and robotic systems, respectively. This mapping enables one to discriminate features of a biorobot that carry a modeling weight from theoretically unconstrained “implementation details.” In addition to this, calibration experiments are often needed to tune robot behavior so as to meet mechanism schema requirements or else to achieve working implementations from underspecified mechanism schemata. And experimental environments have to be carefully monitored for disturbing factors that affect both modeling features and implementation details of biorobots.

The assumptions that are gradually introduced in the process of setting up and performing biorobotic experiments become integral parts of the background hypotheses that are needed to bring experimental observations to bear on biological mechanism descriptions. Distinctive background assumptions in biorobotics concern implementation accuracy with respect to mechanism schemata and descriptions. Genuine examples of biorobotic research are selectively introduced here to illustrate the methodological complexities arising in the preparation of “good” biorobotic experiments or the evaluation of their impact on empirical hypotheses about biological sensorimotor mechanisms.

2. Theoretical Models and Biorobotic Experiments

We take here mechanisms to be “entities and activities organized such that they are productive of regular changes from start or set-up to finish or terminating conditions” (Machamer, Darden, and Craver Reference Machamer, Darden and Craver2000, 3).Footnote ² Accordingly, mechanism descriptions include both temporal and spatial constraints. Temporal constraints concern order, rates, duration, and frequency of activities; spatial constraints concern structure and orientation of entities, in addition to compartmentalization, localization, and interconnection of activities (Craver and Darden Reference Craver, Darden, Machamer, Grush and McLaughlin2001). Mechanism descriptions can be organized into multi-level hierarchies, so that some mechanisms figure as modular components of higher-level mechanisms. Mechanism schemata are more sparing than mechanism descriptions, insofar as no reference to specific material systems is made there. Role functions mentioned in mechanism schemata can be “filled with descriptions of known component parts and activities.” In particular, “mechanism schemata can be instantiated in biological wet-ware … or represented in the hardware of a machine” (Machamer, Darden, and Craver Reference Machamer, Darden and Craver2000, 17).Footnote ³

The discovery of biological sensorimotor mechanisms starts with the characterization of the phenomenon of interest (Craver and Darden Reference Craver, Darden, Machamer, Grush and McLaughlin2001, 121). And the initial stages in the discovery of biological mechanisms often involve the formulation of mechanism sketches, which allow for functional gaps (Machamer, Darden, and Craver Reference Machamer, Darden and Craver2000; Craver and Darden Reference Craver, Darden, Machamer, Grush and McLaughlin2001) or imprecise specification of functional roles. Mechanistic sketches are progressively worked out and revised until a mechanism description is obtained. Underspecification of mechanism sketches is the source of arbitrary choices in biorobot design. In order to set up meaningful biorobotic experiments, one has to understand the contribution of these arbitrary choices on the shaping of overall biorobot behavior and to preserve implementation accuracy with respect to mechanism sketches or schemata.

Mechanism schemata may admit a variety of machine instantiations and implementations. Computer simulation of mechanism schemata, for example, is pursued to support both prediction and explanation (Amit Reference Amit1998) from mechanism descriptions insofar as, by running the simulation, one achieves a seamless account of how “termination conditions are produced by the set-up conditions and intermediate stages” (Machamer, Darden, and Craver Reference Machamer, Darden and Craver2000, 3), one identifies some component contribution to the regular functioning of the overall mechanism, and predicts what behavioral changes manipulative interventions give rise to in accordance with the regularities regimenting interactions between mechanism components (Woodward Reference Woodward2002).

Biorobots are built starting from mechanism schemata or sketches to support scientists much in the way that computer simulations do. Robotic implementations, however, enable one to sidestep major modeling problems that arise in computer simulations of mechanism schemata. In a computer simulation, one has to represent environmental factors that release, sustain or modulate activities of mechanism components. In contrast with this, the need for environmental simulation does not arise in biorobotic investigations, insofar as robotic and biological systems are behaviorally compared to each other in real world environments (typically in the ecological niche of the biological system under investigation). Thus, biorobotic experimental practice does not give rise to the methodological problem of controlling whether behavioral (dis-)similarities between target biological systems and computer-simulated agents take their origin in ad hoc or inaccurate simulations of the environment. One should be careful to note, however, that there is a methodological trade-off between computer simulation and robotic implementation of mechanism schemata. Computer simulations are unaffected by methodological problems arising in biorobotics on account of the fact that biorobots are immersed in natural environments. Indeed, biorobot malfunctioning may be induced by environmental features (temperature, humidity, illumination, and so on) that are perfectly normal for the biological system under investigation. Accordingly, biorobotic experimental settings have to be monitored for two different sets of boundary conditions, with the aim of preserving “normal” functioning conditions for both biological and robotic systems there.

These preliminary observations are sufficient to bring out some distinctive features of biorobotic approaches to the investigation of biological sensorimotor mechanisms. The implementation of a biorobot presupposes that the role functions mentioned in some mechanism sketch or schema M admit both a biological instantiation $M_{S}$ and a robotic instantiation $M_{R}$ . The robotic instance $M_{R}$ , which makes reference to robotic parts and activities, is used to guide the implementation of some robotic system R. Thus, M is the theoretical basis licensing similarity judgments between $M_{R}$ , a biological instance of M, and the implementations of both. Experiments involve behavioral comparisons between R and a biological system S. The latter is hypothesized to implement some biological instance of M. The background assumptions that are needed to bring these experiments to bear on this hypothesis include specific biorobotic assumptions about R's design and implementation process.

There are various ways of bringing biorobotic experiments to bear on mechanism descriptions and sketches. In addition to providing data for model corroboration or falsification, these experiments may enable one to identify previously undetected functional components in the target biological system, to identify causally relevant features of the environment, and to suggest specific expansions of mechanism sketches. Biorobotic experiments, which invariably involve running a biorobot and observing its behavior, may differ from each other along more specific dimensions. One can observe overt motor behavior (such as the robot trajectory), the behavior of internal components (such as the firing rate of artificial neurons embedded in the robot control system), or environmental properties (such as the spatial distribution of chemical gradients to which the robotic system reacts or physical modifications of the environment induced by robot actions). In preparing an experiment, one can manipulate both environmental conditions (by occluding, say, sensory cues or modifying illumination conditions) and internal variables (by modifying, say, parameters in system modules). Thus, biorobotic experiments afford the opportunity of comparing robotic and biological system behaviors in varying external or internal conditions. Distinctive methodological problems this experimental practice gives rise to are analyzed in the following sections, chiefly by reference to biorobotic inquiries on adaptive rat navigation and lobster chemotaxis.

3. Initial Corroborations and New Biological Mechanism Predictions

The capability of rats to learn the spatial structure of mazes, and to reach previously encountered goal locations, such as food sites, has been extensively investigated in the past decades. In the wake of Tolman (Reference Tolman1948), O'Keefe and Nadel (Reference O'Keefe and Nadel1978) proposed to explain these capabilities by reference to some sort of internal map of the environment which is acquired by learning. And empirical evidence emerged for the hypothesis that the hippocampus—a bilateral structure placed between the cerebral cortex and the thalamus—plays a key functional role in spatial tasks. Notably, rats with hippocampal lesions are unable to perform a variety of spatial tasks (Barnes Reference Barnes1988; Jarrad Reference Jarrad1993); and neurons in areas CA3 and CA1 of the hippocampus of freely moving rats—called “place cells”—have been found to fire maximally when rats reach specific locations of the testing platform (O'Keefe and Dostrovsky Reference O'Keefe and Dostrovsky1971; Fox and Ranck Reference Fox and Ranck1975). On account of this finding, place cells were hypothesized to support rat self-localization and heading capabilities.Footnote ⁴

Empirical law-like correlations between place cells firing and rat position enable one to predict rat position on the basis of place cells firing (Wilson and McNaughton Reference Wilson and McNaughton1993). But in order to provide a mechanistic explanation of rat spatial navigation which takes into account place cells behavior, one has to regard place cells as elements of a complex of interacting components which comprise perceptual systems and effectors. How is it that place cells firing drive rat movements? How is it that sensory stimuli determine place cells firing? And how are hippocampal maps acquired by learning? These questions point to crucial gaps in our understanding of the mechanisms producing rat navigation behaviors and the role of place cells therein. In order to fill in some of these gaps (Burgess et al. Reference Burgess1997), Burgess et al. (Reference Burgess2000) proposed a more comprehensive mechanistic model (let us call it $M_{S}$ ) which we turn to summarize now.

$M_{S}$ postulates components formed by various neural maps with Hebbian connections, that are modifiable by unsupervised learning processes. According to $M_{S}$ , a ‘sensory’ map is organized into several ‘sensory cells groups’, each one of them encoding distances from one of the walls surrounding the environment. An ‘entorhinal’ map encodes the distance between the rat and two adjacent walls (each entorhinal cell being connected with two sensory cells of distinct groups). Entorhinal cells are connected with place cells. The receptive field of the former is sharper than the receptive field of the latter.Footnote ⁵ The place cell layer is connected with a ‘goal cell’ layer. Each goal location is represented there by a set of goal cells; and each element of this set is tuned so as to respond maximally when the system is displaced from the goal into a specific direction. Assuming that four goal cells only encode for a goal location, GC_N fires when the rat is north of the goal, GC_S fires when the rat is south of the goal, and so on. The firing rate of each goal cell is proportional to the displacement of the system from the goal in the corresponding direction. The displacement and the heading of the system is given by the vector sum of the directions represented by each goal cell, weighted by their firing rate. Learning proceeds in a feedforward fashion: while the rat is wandering, an unsupervised competitive learning algorithm modifies the connections between entorhinal and place cells, thus creating a place cell map of the environment; one-shot strengthening of synaptic connections between goal cells, ‘head direction cells’ (signaling system heading), and active place cells obtains whenever a reward signal (indicating the presence of the goal) is received. After learning, the firing of goal cells will depend on rat position (signaled by active place cells) and heading. Movements are issued in the directions coded by goal cell vector representations.

This mechanistic model fills in gaps of mechanistic sketches which make reference to place cells and their functional roles. In particular, this mechanism description specifies subsystems which generate heading (on the basis of place cells outputs), determine place cell firing (on the basis of sensory cell representation of distances from walls), and provide an account of hippocampal map learning. Burgess and O'Keefe tested this mechanism for accurate learning of the environment and successful navigation towards previously learnt goals by robotic simulation experiments. These were carried out on a robotic system R consisting of a small wheeled platform, with an on-board camera, infrared proximity detectors, and a neural network which includes 60 sensory cells, 900 entorhinal cells, 900 place cells, and four goal cells representing a single goal position. The experimental environment is a rectangular arena formed by white walls and a dark floor, the north wall being signaled by a distinctive mark. At each processing cycle, the robot rotates in the estimated north, south, east, and west directions, and captures images of the four walls; these images are processed by a black and white edge detection algorithm, which reveals corners between walls and floor; corners are used to evaluate the distance of the system from the walls; by separately controlling the velocities of the two motors, the robot turns in the direction computed by the neural network and moves 3 cm forward.

In experimental trials, the robotic rat learnt successfully an accurate map of the environment, and was able to reach previously learnt goal locations. Initially, the robot was set free to explore (by wandering) a 50×50 cm square environment, according to the mechanism described above. Robot performance after exploration was evaluated in the same environment, and in a “stretched” 50×75 cm environment. In additional trials, the activity of artificial neurons was recorded during navigation in the stretched environment. While entorhinal cells maintained their peaks at fixed distances from the walls, the receptive field of some place cells became stretched and occasionally two distinct peaks were recorded. Notably, this behavior of artificial place cells in stretched environments was found to match the behavior of biological place cells in similar conditions (O'Keefe and Burgess Reference O'Keefe and Burgess1996).

The behavioral match between actual rats S and robotic rats R in similar environments was evaluated chiefly in terms of goal reaching.Footnote ⁶ The experimental outcomes were taken to provide initial corroboration for $M_{S}$ and the mechanistic explanation of goal reaching by rats which flows from it (Burgess et al. Reference Burgess2000). Thus, data about robotic behavior were used as evidence for a theoretical hypothesis about a biological mechanism. Clearly, this corroboration claim, which is made on the basis of an observed behavioral match between R and S, presupposes that robotic rat R is an accurate implementation of mechanism description $M_{R}$ and that $M_{R}$ is an instance of some abstract mechanistic schema M which admits $M_{S}$ as biological instance.

Predictions which $M_{S}$ makes on the receptive fields of sensory, entorhinal, place, and head cells have been corroborated by purely biological experiments. It is worth noting, in this respect, that biorobotic inquiries can motivate biological experiments for identifying biological entities and activities postulated by some mechanistic model. This additional role for biorobotic inquiry is exemplified by reference to the ‘goal cell’ layer of $M_{S}$ . The robotic analog of this layer accounts for the capability of robotic rats to navigate towards goal sites that are not signaled by sensory cues. In additional biorobotic measurements performed in the stretched environment, the goal cells representation was found to be stretched as well, and to affect navigation on this account: “the principal effect of expanding an environment is a separation of the peaks of each goal cell's firing-rate map (the locus of search remaining between them). By contrast, contracting an environment by a large enough factor can cause the locations of peak firing of opposing goal cells to cross over, and produces a more dramatic effect: the robot searches only at the edges of the environment” (Burgess et al. Reference Burgess1997, 1541). These predictions on the behavior of goal cells may guide the search for cells exhibiting a similar behavior in the rat nervous system; and the ‘goal cells’ hypothesis may be tested by observing search patterns of rats in stretched environments, and comparing them to predictions obtained by means of biorobotic systems.

In concluding this section, let us note that new components and role functions of rat navigation mechanisms have been recently hypothesized on account of the discovery of grid cells in the dorsocaudal medial entorhinal cortex of rats. Grid cells might support some spatial computations that were previously attributed to the hippocampus (Hafting et al. Reference Hafting2005). This recent discovery does not affect the broad methodological picture outlined in this section on the basis of biorobotic inquiries on place cells, insofar as biorobotic experiments may prove useful to fill in the gaps of mechanism sketches which include grid cells, to test the behavioral predictions of the resulting mechanism schemata, and to drive the search for biological components and activities interacting with grid and place cells.

4. Feedback-Based Hypotheses on Lobster Chemotaxis

Biorobotic experiments were performed by Grasso et al. (Reference Grasso2000) to investigate the capability of lobsters to track turbulent chemical plumes to their sources (chemotaxis). The mechanism investigated there is based on feedback analysis: lobster antennae can detect local gradients in the chemical plume; and local gradients provide lobsters with directional information to localize the plume source. Computer simulations appear to be unsuitable to test mechanism schemata that are consistent with this broad picture of lobster chemotaxis, insofar as accurate replication of plume turbulence in water is unfeasible (especially in view of the difficulty of developing accurate fluid-dynamical models at the temporal and spatial scales that are needed to simulate lobster sensorimotor loops in chemo-orientation tasks). In contrast with this, robots placed in real chemical plumes allow for direct comparison between robotic and biological behaviors.

Two feedback-based mechanisms have been implemented in a small mobile wheeled robotic platform, called RoboLobster, which is provided with two antennae for chemical gradient detection. By operating the first mechanism (M1), when gradient exceeds a threshold the system turns towards the side of the sensor detecting higher concentration or moves forward when same concentration information is detected by both sensors. The second mechanism (M2) is obtained from M1 by adding a rule, which causes the system to back up when the concentration level detected by both sensors falls below a threshold and until the threshold is exceeded again. In experimental trials, RoboLobster is immersed in a stream of water to which a turbulent plume was added. The source of the plume is immersed too, and the plume assumes a typical oblong shape (replicating the hydrodynamic output from a filter-feeding bivalve, which is a natural food item for lobsters). The plume shape is easily replicated in various trials. Steering is achieved by independently controlling the two wheels.

Several features of RoboLobster motor behavior are monitored while the robot is searching for the plume source. These include closest approach (the shortest observed distance between robot and source) and path tortuosity (the angular variance of turns measured over the entire path). In trials running M1 RoboLobster consistently failed to hit the source (the closest approach distance was high). Better results were obtained by M2, except when the robot started 100 cm away from the source. The authors called ‘distal patch field’ (DPF) this region of the plume, in which both algorithms performed poorly. Before claiming falsification of both mechanism schemata, the authors of this biorobotic study were careful to test whether the selected distance between antennae could affect RoboLobster performance, due to the impossibility of detecting the gradient between close points in the DPF region. It turned out, however, that changing the distance between antennae does not improve RoboLobster performance in DPF. Additional trials were performed by systematically changing initial orientations of the robot, to control whether successful hits of the source from the DPF were due to particular starting positions, but no significant effect of starting orientation on robot performances was found.

Having excluded distance between antennae and starting orientation as alternative explanations for observed behavioral discrepancies, these biorobotic experiments were taken to show that in the DPF, no matter how initial conditions change, neither mechanism accounts for target system behavior. In particular, these results were directed against a specific assumption embedded in M1 and M2, which concerns the structure of the plume in the DPF. The fact that feedback analysis alone is inadequate to account for chemotaxis in the DPF is taken to suggest that “there is no concentration gradient information available to either algorithm in the DPF” (Grasso et al. Reference Grasso2000, 126). A local gradient is instead detectable in regions that are nearer to the plume source, thus allowing for effective robotic chemotaxis.

To sum up: The RoboLobster experiments show that biorobotics may enable one to falsify mechanism schemata and sketches, and to isolate significant, but previously overlooked features of the environment. (“In one sense the plume, rather than the algorithms or the American lobster, is the entity under the closest scrutiny in these experiments” [Grasso et al. Reference Grasso2000, 127].) In addition to this, the RoboLobster inquiry illustrates how biorobotic experiments performed after selective manipulations of environmental or internal features contribute to rule out specific disturbing factors or the influence of theoretically unconstrained options in experimental settings. These experimental strategies in biorobotic inquiry supplement those illustrated by reference to rat navigation mechanisms, whereby biorobotic experiments may enable one to isolate mechanism schemata accounting for biological behaviors, and drive the search for biological entities and activities producing target system behaviors.

5. Biorobotics and the Upgrading of Implementation Details

The discovery of mechanisms often proceeds by piecemeal elaboration and revision of mechanistic sketches (Craver and Darden Reference Craver, Darden, Machamer, Grush and McLaughlin2001, 133–134). Biorobotics can assist the iterative specification of mechanism descriptions from mechanism sketches. One starts with a sketch representing an initial ‘guess’ on target system mechanisms, and progressively identifies role functions and fillers for gaps or underspecified components of the original sketch. Sensory processing and motor control of rat navigation are underspecified components in the mechanistic hypothesis proposed by Burgess and O'Keefe. Their model assumes that the distance between the system and the walls is encoded in sensory cells, and that the goal cell layer outputs a vector pointing to the goal. However, the model specifies neither the mechanistic connections between sensory cells and sensory receptors, nor the connections between goal layer and motor actuators. In particular, the model does not specify how it is that visual, auditory, olfactory, and other sensory stimuli are processed so as to determine sensory cell firing. Similarly, the model does not specify how rat muscles are activated on the basis of the vector computed by the goal cell layer. To continue, this model of rat navigation does not provide an estimate for the population of sensory cells that represent distances (this parameter is crucial to determine how coarse the encodings of distance are). And unfixed parameters are involved in the modeling of sensory neurons too.

The filling in of these mechanistic gaps is achieved in the robotic rat implementation by adding mechanisms for sensory processing and motor control that are unconstrained by the initial hypothesis. In particular, the robot is provided with an artificial camera and with infrared proximity detectors as sensors, and with two independently driven wheels to move in the environment. And the mechanism used to estimate distances from the walls is based on the extraction of the dark-white corner between the wall and the floor, as well as on geometric calculations outputting the distance and the orientation of the wall with respect to the system.

Success in replicating biological behaviors may suggest that some role functions filled in by theoretically unconstrained implementation details be included in the biological hypothesis. For example, experiments on the robotic rat may lead one to conjecture that the rat nervous system exploits light-dark contrast to evaluate distances from the walls. This hypothesis may be tested by observing the behavior of rats in poorly contrasted arenas and seeing whether their performance decreases there; and one may additionally run biorobotic experiments with a modified robotic rat using, say, an alternative visual processing algorithm and checking whether the behavioral match between biological system and robot improves on account of this modification. Indeed, alternative goal-seeking mechanisms were implemented, tested, and the selection between them was performed by Burgess et al. (Reference Burgess2000, 305) on the basis of some behavioral match criteria.

Similar considerations arise in connection with theoretically unconstrained implementation features that are added for the sole purpose of preserving normal working of the robotic system. Suppose that mechanism description M postulates an open-loop control of actuators which is consistent with neural recordings of motoneurons; and suppose that feedback control is required in robotic implementation of M insofar as, say, rough surfaces in the experimental ground hinder robot movements and give rise to significant discrepancies between desired and actual movements. Corrective motor commands can be issued by endowing the system with proprioceptive sensors (e.g., a shaft encoder which detects the actual steering of the system) and with a feedback-based mechanism. If the robotic system including this feedback mechanism—which is initially not postulated in M—successfully replicates biological behaviors, then one may look for a similar mechanism in the biological system, which includes proprioceptive sensors and feedback-based inhibiting neurons, and counteracts the effects of rough surfaces on biological system movement. This experimental possibility is an additional example of a broad biorobotic strategy for the iterative specification of mechanism descriptions: the role functions performed by arbitrary implementation details of biorobots are selectively introduced in mechanism sketches and schemata on account of the successful replication of biological behaviors. The upgraded model, in its turn, prompts the search for independent evidence about biological entities and activities supporting these role functions.

6. Background Assumptions in Biorobotic Experiments

The need for distinctive background assumptions has been diffusely emphasized in the above discussion of biorobotic experiments. In addition to multiple instantiation hypotheses for mechanism schemata and normalcy conditions for robot behavior, these assumptions notably concern implementation accuracy: if the entities and activities of robotic system R do not fulfill the constraints set by mechanism description $M_{R}$ , then one can hardly extract from an analysis of R's behaviors significant information on the relationship between target biological system S and biological mechanism description $M_{S}$ (which is supposed to preserve the role functions of $M_{R}$ ).Footnote ⁷ The significance that implementation accuracy assumes in biorobotics becomes particularly evident by contrasting simulation requirements in biorobotics and biologically-inspired robotics. Analyses of biological solutions to sensorimotor control problems play heuristic roles in biologically-inspired robotics, but accurate modeling of biological mechanisms is not pursued when more efficient solutions to engineering problems are at hand. Consider, for example, biological response delays. Mechanism descriptions of sensorimotor coordination in biological systems take into account neural delays occurring in both sensory data and motor command transmission and processing. Current robotic technology often allows for faster responses. Thus, a close match between biological and robotic mechanism descriptions may require appropriate robotic simulation of neural delays which, in fact, decrease overall robot performance.Footnote ⁸

Accuracy may be compromised in the process of biorobotic implementation by adaptations of theoretically constrained items or additions of theoretically unconstrained implementation details. Let's see.

Adapting theoretically constrained items. Perfect functional mapping of biological mechanism descriptions onto robotic mechanism descriptions is a regulative ideal for biorobotic inquiry, which is hardly ever attained in practice, in view of practical and theoretical impediments. Practical impediments chiefly originate in limitations of current robotic technology, insofar as hardware components meeting every constraint imposed by mechanism descriptions are not available. Consider, for example, sensors for detecting chemical plumes that are needed for chemotaxis simulation purposes. As pointed out in Ishida et al. (Reference Ishida2001, 222), “sensors for a robot to detect those signals with capabilities similar to those of animals are not yet available.” Striking limitations of available sensors concern temporal features of response onset and recovery of initial conditions after stimulus removal (which may amount to 30s for some sensor technology). Thus, opting for the use of some particular sensor technology becomes a theoretically significant decision, to the extent that constraints on sensitivity, response or recovery times of chemical sensors are specified in mechanism descriptions.

Adaptations are occasionally pursued even when adequate modeling hardware is available, in order to facilitate implementation and reduce control problems. Consider, in this connection, the detailed model of hexapod locomotion formulated by Quinn and Ritzmann (Reference Quinn, Ritzmann, Webb and Consi2001), which takes into account insect leg kinematics. Since accurate replication of the insect kinematic chain in robot hardware gives rise to difficult implementation and control problems, some degrees of freedom in robotic legs were provisionally disregarded. Simulation data obtained on the basis of the simplified implementation were encouraging, insofar as they suggested that the eliminated degrees of freedom were immaterial to a satisfactory operation of the middle and rear legs. In general, however, the effects of these adaptations may become unpredictably significant in real-world experimental settings and under uncontrolled boundary conditions. Therefore, the background assumptions arising from such simplifying adaptations have to be isolated and possibly supported by independent experimental results, in order to make proper use of adapted robots in biorobotic experimental inquiries.

Selecting theoretically unconstrained items. The rat navigation mechanism proposed by Burgess and O'Keefe does not constrain the choice of motor actuators in the process of robotic implementation. Viable options, which do not violate mechanism description constraints, include various kinds of independently controlled wheels or leg arrays (endowed with different degrees of freedom). Similar considerations apply to the selection of values for underspecified parameters in mechanism sketches, possibly achieved by calibration, as long as this selection does not lead one to violate mechanism sketch constraints.

Even though no violation of theoretical constraints is introduced in these cases, one should be careful to note that the overall system behavior may depend on options concerning theoretically unconstrained items. Consider, for example, different visual mechanisms for detecting the rat's distance from the walls. Visual processing time is not constrained by the mechanistic hypothesis under investigation, but different processing speeds may affect system behavior in significant respects, including system reactivity to visual stimuli and the achievement of an adequate neural representation of the environment. Behavioral match judgments concerning artificial and biological behaviors have to take into account (and possibly factor out) the effects of these arbitrary implementation choices. Similarly, different values for theoretically unconstrained parameters may produce similar behaviors with respect to a limited range of set up conditions, and significantly differing behaviors outside that range. Thus, biorobotics gives rise to special cases of the arbitrary value adjustment problem affecting mechanism-based modeling in biology at large, informatively discussed by Hopkins and Leipold (Reference Hopkins and Leipold1996).

Additional background assumptions that distinctively arise in biorobotics concern boundary conditions that are special to biorobot behavior. Clearly, alterations of environmental or internal system conditions may reveal limits in the scope of regularities governing the behavior of the target biological system and its components. Similarly, the boundary conditions ensuring “normal” behavior of a biorobot in experimental settings may not coincide with the boundary conditions ensuring normal behavior in the target biological system. For example, a plausible constraint on the proper behavior of many robotic systems is that environmental temperature does not exceed 40°C, insofar as the performance of microchips rapidly deteriorates at higher temperatures. This condition concerns a theoretically unconstrained feature of the implemented system; it can be dropped if one replaces microchips with more robust chips or software code (a similar issue is dealt with in Reeve et al. [Reference Reeve2005]). This example illustrates from another perspective the dependence of biorobot behaviors on a wide variety of theoretically unconstrained details. Failure of the robotic rat to perform properly in goal-seeking tasks may be attributed to disturbing environmental conditions of this sort. Accordingly, bioroboticists have to monitor experimental environments for boundary conditions concerning robot material components, and particularly items that are theoretically unconstrained in the implementation, in order to make sensible attributions of behavioral mismatches to incorrect predictions flowing from mechanism descriptions and schemata.Footnote ⁹

7. Summary

Biorobotic experiments involve running robotic implementations of mechanism schemata and descriptions, and evaluating their behavior with respect to the behaviors produced by the target biological system in similar experimental environments. Behavioral match evaluations of robotic and biological behaviors may enable one to corroborate or falsify mechanism schemata and descriptions, to identify previously overlooked interactions with the environment, to suggest the existence of entities and activities implementing some role function in the target biological system, and more generally to support the iterative specification of mechanism descriptions from mechanism sketches.

These roles for robotic material models in experimental inquiries about adaptive sensorimotor mechanisms have been chiefly explored here by reference to biorobotic experiments concerning rat navigation and lobster chemotaxis behaviors. In the former case, biorobotic experiments were performed which suggest the existence of particular entities and activities in the biological mechanism under investigation, in the light of the fact that a biorobot successfully duplicating target biological system behaviors was obtained by filling in gaps or underspecified components of mechanism sketches and implementing the resulting mechanism description. In the latter case, biorobotic experiments were performed which call and heuristically guide the search for a new mechanistic model, insofar as these experiments show that the mechanism description instantiated and implemented in a biorobot fails to produce specific behaviors of the target biological system.

Biorobotic experiments enable one to address a variety of additional questions about adaptive sensorimotor mechanisms. If the set of variables included in some mechanism description is reasonably small, one may use biorobotic experiments to look for new correlations in a systematic fashion (see, e.g., Chou and Hannaford Reference Chou and Hannaford1997; Jaax and Hannaford Reference Jaax and Hannaford2002). And minor modifications of a mechanism description accounting for, say, rat navigation behavior may enable one to replicate in a biorobotic experiment adaptive navigation by another biological species, thus suggesting that similar mechanisms produce the navigation behavior of different biological species. Careful examination of these additional experimental roles for biorobots is called for in a thorough methodological analysis of biorobotics.

Mechanistic theorizing based on biorobotic experiments rests on a variety of distinctive background assumptions, chiefly concerning implementation accuracy and multiple instantiation of mechanism schemata. In particular, we have emphasized the role of implementation accuracy assumptions in the inference process which leads one to claim corroboration (falsification) for some mechanistic hypothesis about some biological system S, when starting from the experimental observation of matching (mismatching) behaviors of biological system S and a biorobot R.

Implementation accuracy with respect to some mechanistic sketch or description is compromised by implementation choices violating theoretical constraints (either imposed by lack of adequate robotic hardware or introduced to reduce implementation difficulties) and by adaptations of theoretically unconstrained implementation details (achieved, say, by adding modules or calibrating robot parameters). Additional methodological problems that are special to biorobotics, and have to be properly addressed to bring robotic behaviors to bear on biological mechanism schemata and sketches include the selection of adequate matching criteria for comparing behaviors produced by structurally different (biological and robotic) systems and the identification of boundary conditions that affect robot behavior but do not affect biological system behavior in the same experimental setting.

On the whole, biorobotic inquiries pose distinctive methodological problems that do not arise in computer simulation studies or in experiments involving biological systems only. Notably, the methodological problem of identifying undetected environmental features causing robot malfunctioning does not arise in experiments involving computer simulations, insofar as the environment is fully controlled there. And the need for controlling biorobot implementation accuracy does not arise in experiments involving biological systems only. An estimate of the theoretical and experimental efforts that are needed to address these distinctive methodological problems may prove crucial to deciding whether a biorobotic experimental strategy is worth pursuing in the investigation of biological sensorimotor mechanisms.Footnote ¹⁰

8. Concluding Remarks

Relatively recent developments of robotic technology have paved the way to the wide variety of biorobotic experiments motivating the present methodological analysis. Recent advances in bionics are now paving the way to novel experimental uses of robotic systems and components in biological inquiry. In bionic systems, mutually interconnected robotic and biological parts are often interfaced to the central or peripheral nervous system of some biological entity. This was achieved, for example, in a bionic system including a lamprey brain stem (Reger et al. Reference Reger2000), engineered to perform sensorimotor transformations controlling motion of a wheeled robotic platform toward light sources. Prosthetics and rehabilitation arguably provide the primary motivation and range of applications for bionic research (Dario et al. Reference Dario2005). Bionic systems, however, naturally suggest a wide range of experiments related to purely theoretical inquiries into biological sensorimotor mechanisms. Indeed, bionic systems implemented on the basis of abstract mechanism schemata which admit both biological instantiation and “hybrid” bionic instantiation might contribute to discovering and testing the functional roles of biological components in sensorimotor tasks. Consider a purely notional bionic assembly, obtained by substituting nervous tissue extracted from some biological system for the feedback-analyzer module of a light-seeking robot. Experiments involving this bionic mechanism may enable one to isolate the functional contribution of that biological component to the light-seeking behavior of the biological system it was extracted from. Additional experiments involving bionic assemblies may concern the stimulation of learning or adaptation processes in biological nervous systems. For example, the observation of behavioral plasticity in human subjects whose peripheral nervous system is connected to a robotic arm prosthesis might shed light on mechanisms for updating internal models of parts of the body in the cerebellum. Thus, one is naturally led to extend the above methodological analysis towards methodological issues distinctively arising in bionic experimentation about sensorimotor mechanisms.

Stimulation and inhibition experiments, conducted by means of bionic or purely robotic systems, can be deployed to discover or test hypotheses about mechanism decomposition hierarchies, in which lower level mechanism descriptions are obtained from higher level descriptions by opening black boxes and identifying new regularities involving system variables. This decomposition hierarchy, however, has to be carefully distinguished from another kind of hierarchical organization of mechanism schemata and descriptions, which is obtained by operations of abstraction or instantiation. Mechanism schemata admit different instantiations, and each one of these may come in different degrees, insofar as one can provide a partial instantiation of an abstract mechanism schema (by indicating, e.g., appropriate fillers for only a few of its functional roles). Clearly, a mechanism schema M can be instantiated in mechanism description $M_{S}$ without “opening” any functional box in the transition from M to $M_{S}$ (by mere identification of fillers for the boxes that are already specified in M). And M can be decomposed into sub-modules without providing any instantiation for them. The interplay between functional decomposition and instantiation hierarchies is crucial for the methodology of biorobotic experimentation at large. To illustrate, consider the above mentioned methodological problems concerning implementation details. Working robotic systems are obtained by appropriately instantiating abstract schemata and implementing the resulting mechanism descriptions. To achieve this result, one is forced to identify artificial entities and activities “all the way down,” so as to generate proper system behavior. But instantiation and implementation of some mechanism schema M does not coincide with the “opening of boxes” in M, insofar as some entities and activities in the material robotic system are bound to be theoretically unconstrained implementation details. An instantiated mechanism description does not necessarily specify all relevant causal factors; and instantiation of mechanism schemata is not obtained by further modular decomposition. On the whole, biorobotics appears to provide a suitable framework to pursue an analysis of the interplay between abstraction and decomposition hierarchies, of their respective role in the identification of different levels of analysis for mechanism entities and activities, of the unification problems that a multilevel mechanistic analysis gives rise to.

The methodological reflections developed here are confined to mechanistic modeling and experimenting in biorobotics. These reflections contribute to outlining a regulative methodological framework for contemporary biorobotic investigations of sensorimotor mechanisms, without questioning the broad mechanistic assumption that adaptive sensorimotor behavior is produced by the interaction of an array of structured and hierarchically organized subcomponents. But an extension of this reflective work, possibly relinquishing this mechanistic assumption, is naturally suggested by the pervasive role of neural networks and dynamical systems in biorobotics: neural network may come with a nonmodular organization; dynamical systems are often used to describe relations between motor variables, such as legs angles and insect motion (Beer Reference Beer1997), without identifying system components and their causal roles in the generation of behavior. Biorobotics appears to be a particularly suited domain for an in-depth exploration of both potentials and limitations of mechanistic modeling and explanation.