Morphological computation (MC), in the context of embodied (artificial) intelligence, refers to processes, which are conducted by the body (and environment) that otherwise would have to be performed by the brain (Pfeifer and Bongard, 2006). A nice example of MC is given by Wootton (1992) (see p. 188), who describes how “active muscular forces cannot entirely control the wing shape in flight. They can interact dynamically with the aerodynamic and inertial forces that the wings experience and with the wing’s own elasticity; the instantaneous results of these interactions are essentially determined by the architecture of the wing itself […]”

MC is relevant in the study of biological and robotic systems. To illustrate the utility of MC in robotics, we will first briefly discuss the Passive Dynamic Walker (McGeer, 1990), which is a purely mechanical system. The proportions of the legs and the weight distribution are similar to those of human legs. The Passive Dynamic Walker shows that walking can result from the interaction of the system’s physical properties and its environment (in this case a slope of 3°). No actuation is required for the walking behavior. Transferred to robotics, this means that systems with high morphological computation only need to generate motor commands when they are needed. Not only does such a control increase the durability of the systems (because the wear-out of the actuators is decreased), it also means that robots with high MC will have a reduced energy demand for their actuation (e.g., Niiyama et al. (2012) and Renjewski et al. (2015)). By outsourcing computation to the embodiment, the controller complexity, and hence, the computational cost can be reduced, which again contributes to the overall reduction of the energy consumption. For contemporary mobile robots, the energy supply still is one of the biggest unsolved problems, as mobile robots can either operate only for a short period of time or have to carry heavy batteries (Baughman, 2005). To summarize, increasing MC for a robotic system potentially decreases the overall energy demand, the controller complexity, and finally also the wear-out of the system.

This said, one should not mistake MC to be synonymous for energy efficiency. One example in which MC does not relate to energy efficiency is human running, in particular, running in an outdoor environment, such as a downhill off-road path. The runner is not able to detect every change of slope or see every stone, tree branch, etc., on the ground. This means that most irregularities are sensed at the moment when the foot touches the ground (Müller et al., 2015). However, an immediate reaction in the sense of neural change in muscle activity is not possible, as the communication pathway from the feet to the spinal cord and back is simply too long (~0.02 s). Running on an irregular ground is only possible because of the physical properties of the muscle-tendon system. Not only do they have a shock-absorbing function, within certain limits, irregularities of the ground can be accounted for without any explicit control (Brown et al., 1995; Proctor and Holmes, 2010; Müller et al., 2015). In this case, MC does not contribute to the energy efficiency (running is not energy efficient), but it is required to enable the behavior. We believe that maximizing MC covers more than just energy efficiency (fast adaptivity in uncertain environments is one example), but that this is an immediate benefit in the field of robotics.

For biological systems, energy efficiency and adaptivity are important and evolutionary advantages. A strong indication for the importance of energy efficiency is given by the fact that the human brain accounts for only 2% of the body mass but is responsible for 20% of the entire energy consumption (Clark and Sokoloff, 1999). The energy consumption is also remarkably constant (Sokoloff et al., 1955). Under the assumption that the acquisition of energy was not a trivial task throughout most of the evolution of humans, on can safely conclude that as much computation as possible has been outsourced to the embodiment. When hunting prey or escaping from a predator, adaptivity to irregularities in the environment through morphological properties is a clear evolutionary advantage. Therefore, MC may be a driving force in evolution.

In biological systems, movements are typically generated by muscles. Several simulation studies have shown that the muscles’ typical non-linear contraction dynamics can be exploited in movement generation with very simple control strategies (Schmitt and Haeufle, 2015). Muscles improve movement stability in comparison to torque driven models (van Soest and Bobbert, 1993) or simplified linearized muscle models [for an overview see (Haeufle et al., 2012)]. Muscles also reduce the influence of the controller on the actual kinematics (they can act as a low-pass filter). This means that the hopping kinematics of the system is more predetermined with non-linear muscle characteristics than with simplified linear muscle characteristics (Haeufle et al., 2012). And finally, in hopping movements, muscles reduce the control effort (amount of information required to control the movement) by a factor of approximately 20 in comparison to a DC-motor driven movement (Haeufle et al., 2014).

In view of these results, we expect that MC plays an important role in the control of muscle driven movement. To study this quantitatively, a suitable measure for MC is required. There are several approaches to formalize MC. Hauser et al. (2011) proposes a dynamical systems approach in which the effect of the morphology on the attractor landscape of the systems is taken into account. Polani (2011) minimizes the controller complexity in a reinforcement learning setting. This is very close to our second quantification MC_MI (see Section 3.2). The difference is that we take the behavior complexity instead of a reward signal into account. Rückert and Neumann (2013) measure morphological computation indirectly by measuring how changes of the morphology reduce or increase the required controller complexity in the context of stochastic optimal control.

In our previous work, we have focused on a direct quantification of the embodiment, whereas most other approaches quantify MC indirectly through the controller complexity. In particular, in our first publication (Zahedi and Ay, 2013), we focused on an agent-centric perspective of measuring MC. We had proposed two different, complementary concepts of measuring morphological computation (which are discussed below). From those concepts, which rely on information that is only partially available to the agent, we derived quantifications that only rely on the agents sensors and actuators. In this publication, we evaluate MC on simulated muscles models, which means that we have access to the full system’s state. Hence, in this publication, we concentrate on measures, which operate on the full system’s state. This allows us to investigate how the concepts operate on realistic muscle models without taking approximation errors into account that occur from limited sensor information. In a different publication (Ghazi-Zahedi and Rauh, 2015), we applied an information decomposition (Bertschinger et al., 2014) to the sensorimotor loop and developed a refined understand the two concepts that we have proposed in Zahedi and Ay (2013). Both publications used a binary toy world model to evaluate the measures. With this toy model, it was possible to show that these measures capture the conceptual idea of MC and, in consequence, that they are candidates to measure MC in more complex and more realistic systems.

The main contribution of this publication is to evaluate two measures of MC on biologically realistic hopping models. With this, we want to demonstrate their applicability in non-trivial, realistic scenarios. Based on our previous findings (see above), we hypothesize that MC is higher in hopping movements driven by a non-linear muscle compared to those driven by a simplified linear muscle or a DC-motor. Furthermore, our experiments show that a state-dependent analysis of MC for the different models leads to insights, which cannot be gained from the averaged measures alone. Finally, we provide detailed instructions on how to apply these measures to robotic systems and to computer simulations, including MATLAB^®, C++ code, and the data used in this publication (Ghazi-Zahedi, 2016). With this, we hope to provide a tool for the evaluation of MC in a large variety of applications.

The quantifications of MC require a formal representation of the sensorimotor loop (see Figure 1), which is introduced in the next section as far as it is required to understand the remainder of this work. For further information, the reader is referred to Klyubin et al. (2004), Zahedi et al. (2010), Ay and Zahedi (2013), and Zahedi and Ay (2013).

The conceptual idea of the sensorimotor loop is similar to the basic control loop systematics, which is the basis of robotics and also of computer simulations of human movement. In our understanding, a cognitive system consists of a brain or controller, which sends signals to the system’s actuators, thereby affecting the system’s environment. We prefer to capture the body and environment in a single random variable named world. This is consistent with other concepts of agent-environment distinctions. An example for such a distinction can be found in the context of reinforcement learning, where the environment (world) is everything that cannot be changed arbitrarily by the agent (Sutton and Barto, 1998). A more thorough discussion of the brain-body-environment distinction can be found in (von Uexkuell, 1934; Clark, 1996) and more recently by Ay and Löhr (2015). A brief example of a world, based on a robot simulation, is given below. The loop is closed as the system acquires information about its internal state (e.g., current pose) and its world through its sensors.

For simplicity, we only discuss the sensorimotor loop for reactive systems in this work (for a detailed discussion, please see Ay and Zahedi (2014)). This is plausible, because behaviors which exploit the embodiment (e.g., walking, swimming, flying) are typically reactive. A measure that takes the controller state into account is discussed in Zahedi and Ay (2013). Restricting ourselves to reactive systems results in three (stochastic) processes S(t), A(t), and W (t), t∈N that constitute the sensorimotor loop (see Figure 1), which take values s, a, and w, in the sensor, actuator, and world state spaces (their respective domains will be clear from the context). The directed edges (see Figure 1) reflect causal dependencies between these random variables. We consider time to be discrete, i.e., t∈N and are interested in what happens in a single time step. Therefore, we use the following notation. Random variables without any time index refer to some fixed time t and primed variables to time t + 1, i.e., the two variables S, S′ refer to S(t) and S(t + 1).

Starting with the initial distribution over world states, denoted by p(w), the sensorimotor loop for reactive systems is given by three conditional probability distributions, β, α, π, also referred to as kernels. The sensor kernel, which determines how the agent perceives the world, is denoted by β(s|w), the agent’s controller or policy is denoted by π(a|s), and finally, the world dynamics kernel is denoted by α(w′|w,a).

To understand the function of the world dynamics kernel α(w′|w,a), it is useful to think of a robotic simulation. In this scenario, the world state W is the state of the simulator at a given time step, which includes the pose of all objects, their velocities, applied forces, etc. The actuator state A is the value that the controller passes on to the physics engine prior to the next physics update. Hence, the world dynamics kernel α(w′|w,a) is closely related to the forward model that is known in the context of robotics and biomechanics.

Based on this notation, we can now formulate quantifications of MC in the next section.

In the introduction, we stated that MC relates to the computation that the body (and environment) performs that otherwise would have to be conducted by the controller (or brain). This means that we want to measure the extent to which the system’s behavior is the result of the world dynamics (i.e., the body’s internal dynamics and it’s interaction with its world) and how much of the behavior is determined by the policy π (see Figure 1). In our previous publication (Zahedi and Ay, 2013), we have derived two complementary concepts to quantify MC (see Section 1 for a discussion). The two measures discussed below were chosen for presentation in this work, because they represent the two concepts.

In a reduced model, hopping motions can be described by a one-dimensional differential equation (Haeufle et al., 2010): (3)my¨=−mg+0y>l0flight phaseFLy≤l0ground contact, where the point mass m = 80 kg represents the total mass of the hopper, which is accelerated by the gravitational force (g = –9.81 m/s²) in negative y-direction. An opposing leg force F_L in positive y-direction can act only during ground contact (y ≤ l₀ = 1 m). Hopping motions are then characterized by alternating flight and stance phases. For this manuscript, we investigated three different models for the leg force (Figure 2). All models have in common, that the leg force depends on a control signal u(t) and the system state y(t), ẏ(t): FL=FL(u(t),y(t),ẏ(t)), meaning, that the force modulation partially depends on the controller output u(t) and partially on the dynamic characteristics, or material properties of the actuator. The control parameters of all three models were adjusted to generate the same periodic hopping height of max (y(t)) = 1.070 m. All models were implemented in MATLAB^® Simulink™ (Ver2014b) and solved with ode45 Dormand-Prince variable time step solver with absolute and relative tolerances of 10⁻¹². To evaluate and compare the results of the models, a time-discrete output with constant sampling frequency is required (see Section 3.3). Therefore, the quasi-continuous variable time step signals generated by the ode45 solver are not adequate. To generate the desired output at 1 kHz sampling frequency, we used the Simulink built-in feature to generate desired output only. This way, the solver decreased step-size below 1 ms if required for precision, however, Simulink would only output the data for 1 kHz sampling frequency. This is similar to measuring a continuous physical system with a discrete time sensor. The models were solved for T = 8 s.

This section discusses the experiments that were conducted with the hopping models and the preprocessing of the data. Algorithms for the calculations are provided in the previous section (Section 3). At this stage, the measures operate on discrete state spaces [see equations (1)–(4) and Algorithms 2–4]. Hence, the data was discretized in the following way. To ensure the comparability of the results, the domain (range of values) for each variable (e.g., the position y) was calculated over all hopping models. Then, the data of each variable was discretized into 300 values (bins). The algorithm for the discretization is described in Algorithm 1. Different binning resolutions were evaluated, and the most stable results were found for more than 100 bins. Finding the optimal binning resolution is a problem of itself and beyond the scope of this work. In practice, however, a reasonable binning can be found by increasing the binning until further increase has little influence on the outcome of the measures.

The possible range of actuator values is different for the motor and muscle models. For the muscle models, the values are in the unit interval, i.e., a(t) ∈ [0,1], whereas the values for the motor can have higher values (see above). Hence, to ensure comparability, we normalized the actions of the motor to the unit interval before they were discretized.

The hopping models are deterministic, which means that only a few hopping cycles are necessary to estimate the required probability distributions. To ensure comparability of the results, we parameterized the hopping models to achieve the same hopping height.

The following paragraphs discuss the findings based on the averaged results presented in Table 1 first and then follow with a discussion of the state-dependent results presented in Figure 3.

All three models generated a similar movement, i.e., periodic hopping with a hopping height of 1.07 m. However, the control signals and the trajectories of the center of mass vary between models (Figure 4). Therefore, the computed values for the morphological computation (MC) vary between the models for both quantification methods (Table 1). Compared to the complex muscle fiber model (MusFib), the linearized muscle model (MusLin) and the technical DC-motor model (DCMot) result in significantly lower values of MC (≈30% less, see Table 1).

To analyze the differences between the models in more detail, we plotted the state-dependent MC (see Algorithm 3). Figure 3 shows the values of MC_W for each state of the models during two hopping cycles. We chose to discuss MC_W only, because the corresponding values of MC_MI are very similar to those of MC_W, and hence, a discussion of the state-dependent MC_MI will not provide any additional insights. The plots for all models and the entire data are shown in Figure 4.

The orange line shows the state-dependent MC for the linear muscle model (MusLin) and the blue line for the non-linear muscle model (MusFib). The green line shows the state-dependent MC for the motor model (DCMot). In the figure, the lower lines show the position y of the center of mass over time. The PD-controller of the DCMot model is parameterized to follow the trajectory of the MusFib model, which is why the blue and green position plots coincide. The original data is shown in Figure 4. There are basically three phases, which need to be distinguished (indicated by the vertical lines). First, the flight phase, during which the hopper does not touch the ground (position plots are above the red line), second, the deceleration phase, which occurs after landing (position is below the red line but still declining), and finally, the acceleration phase, in which the position is below the red line but increasing.

The first observation is that MC is very similar for all models during most of the flight phase (position above the red line) and that it is proportional to the velocity of the systems. By that we mean that MC decreases when the velocity during flight decreases and increases when the system’s speed (toward the ground) increases. During flight, the behavior of the system is governed only by the interaction of the body (mass, velocity) and the environment (gravity) and not by the actuator models. Also, all actuator control signals are constant during flight. This explains why the values coincide for the three models.

For all models, MC drops as soon as the systems touch the ground. DCMot and MusLin reach their highest values only during the flight phase, which can be expected at least from a motor model that is not designed to exploit MC. The graphs also reveal that the MusLin model shows slightly higher MC around mid-stance phase, compared to the DCMot model. For the non-linear muscle model, the behavior is different. Shortly after touching the ground, the system shows a strong decline of MC, which is followed by a strong incline during the deceleration with the muscle. Contrary to the other two models, the non-linear muscle model MusFib shows the highest values when the muscle is contracted the most (until mid-stance). This is an interesting result, as it shows that the non-linear muscle is capable of showing more MC while the muscle is operating, compared to the flight phase, in which the behavior is only determined by the interaction of the body and environment.

This work presented two different quantifications of morphological computation including algorithms, MATLAB^®, and C++ code to use them (Ghazi-Zahedi, 2016). We demonstrated their applicability in experiments with non-trivial, biologically realistic hopping models and discussed the importance of a state-based analysis of morphological computation. The first quantification, MC_W, measures MC as the conditional mutual information of the world and actuator states. Morphological computation is the additional information that the previous world state W provides about the next world state W′, given that the current actuator state A is known. The second quantification, MC_MI, compares the behavior and controller complexity to determine the amount of MC.

The numerical results of the two quantifications MC_W and MC_MI confirm our hypothesis that the MusFib model should show significantly higher MC, compared to the two other models (MusLin, DCMot). These results complement previous findings showing that the minimum information required to generate hopping is reduced by the material properties of the non-linear muscle fibers compared to the DC-motor driven model (Haeufle et al., 2014). More precisely, the higher MC in the MusFib model can be attributed to the force–velocity characteristics of the muscle fibers. Previous hopping simulations showed that these non-linear contraction dynamics reduce the influence of the controller on the actual hopping kinematics (Haeufle et al., 2010, 2012). This means, that the kinematic trajectory is more predetermined by the material properties as compared to the linearized muscle model MusLin, or the motor model DCMot. Similar effects have been demonstrated for jumping (van Soest and Bobbert, 1993) and walking (Gerritsen et al., 1998; John et al., 2013). In conclusion, this implied that studies on neural control of biological movement should consider the biomechanical characteristics of muscle contraction (Pinter et al., 2012; Schmitt et al., 2013; Dura-Bernal et al., 2016).

We also showed that a state-dependent analysis of MC leads to additional insights. Here, we see that the non-linear muscle model is capable of showing significantly more morphological computation in the stance phase, compared to the flight phase, during which the behavior is only determined by the interaction of the body and environment. This shows that morphological computation is not only behavior dependent but also state-dependent. Future work will include the analysis of additional behaviors, such as walking and running, for which we expect, based on the findings of this work, to see more morphological computation of the non-linear muscle model MusFib.

To summarize the previous paragraphs, in this work we have showed that the results obtained from the two measures MC_W and MC_MI correspond to the intuitive understanding of morphological computation in muscle and DC-motor models. Furthermore, the results are in accordance with previous work on the control complexity of these models (Haeufle et al., 2014). We also showed that both measures show very similar results for deterministic systems. Hence, one might ask if there is a justifiable preference for one of the two measures. To answer this question, it is helpful to set the methods and results of this work in perspective to our previous work.

In our first publication (Zahedi and Ay, 2013), we proposed two concepts for quantifying morphological computation. The first concept MC_A = D(p(w′|w, a) || p(w′|w)) was not used in this work, because it not applicable to deterministic systems. The original concepts were developed for stochastic systems only. Instead, we analyzed a replacement, which is the measure MC_MI. In this work, we showed that MC_MI and MC_W are almost equivalent in the context of deterministic systems. In our second work into account, we applied an information decomposition to the sensorimotor loop (Ghazi-Zahedi and Rauh, 2015), and found that the second concept, MC_W, can be decomposed into the unique information that the current world state W contains about the next world state W′ and the synergistic information about the next world state W′ that is only available if the current world state W and current action A are taken into account simultaneously. If we take all of these results into consideration, we can conclude that there are more arguments in favor of MC_W than MC_MI for the following reasons. (1) The decomposition of MC_W leads to quantities, which nicely reflect intuitive understanding of morphological computation (Ghazi-Zahedi and Rauh, 2015). (2) MC_W and MC_MI lead to almost identical results for deterministic systems (see Appendix B). (3) MC_W can be computed more easily. It requires the joint distribution of only three random variables (W, A, W′), whereas MC_MI requires the joint distributions of p(w′, w) and p(a, s), and hence, also information about the sensor states S. This does not mean that we conclude that MC_W is the final answer to the question on how to quantify morphological computation. Yet, these are strong indications, that the second concept proposed in Zahedi and Ay (2013) is the path to take for further investigations. Nevertheless, the results of this paper show that MC_W can already be used to quantify morphological computation in realistic systems. Hence, the next step is to apply MC_W on high-dimensional data collected from natural systems. The first issue that needs to be solved here is the formulation of MC_W for the continuous domain, which is our next step.

We believe that quantifying morphological computation is also useful in the context of robotics, e.g., in the optimization of morphology and control. Hence, in currently ongoing work, we are applying MC_W (and the provided software, which we hope proves to be beneficial also to others) to study the effect of morphological design and controller variations on morphological computation in the context of soft robotics.

The paper was mainly written by KG-Z and DH but with strong contributions by the other three authors. The experiments were conducted by KG-Z and DH with strong contributions by GM. All authors contributed significantly to the discussion of the results and the final manuscript.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.