Part V — Methodology and advanced topics · Chapter 16

Morphological computation and the evolution of things

~55 min read2 interactive widgets2 plates

In this chapter

  1. The second major transition: evolution in materio
  2. Morphological computation: the offloading debate
  3. Four characteristic cases
  4. Three classes: control, perception, and computation proper
  5. Design implications: soft robotics and the body as a resource
  6. Nanowire networks: a new substrate for control
  7. Online adaptation: behaviour repaired in runtime
  8. Fault recovery with Boolean-network robots
  9. Reconfigurable organisms: from evolution to living machines
  10. Closing the loop: the course's promises, cashed
  11. Check your understanding

1. The second major transition: evolution in materio

Eiben and Smith (2015) frame the history of artificial evolution as two major transitions. In the twentieth century, computer technology enabled artificial Darwinian processes in silico — the evolution of digital entities: evolutionary computation (Chapter 10). In the twenty-first century, developments in robotics, materials science and 3D printing enable the evolution of physical artefacts or machines — in materio: the evolution of things.

TWO MAJOR TRANSITIONS IN ARTIFICIAL EVOLUTION WETWARE in vivo natural evolution (the twentieth century) SOFTWARE in silico evolutionary computation (Ch. 10) HARDWARE in materio evolution of things (this chapter) 1st 2nd "If you have variation, heredity, and selection, then you must get evolution" (Dennett) the substrate changes — the Darwinian principles do not
Plate 16.1 — The two transitions of Eiben & Smith (2015). The evolution of things closes the loop that Chapter 10 opened with the GOLEM project: not only is the design evolved, the physical artefact is evolved and manufactured.

The reasons this matters for robots: for robots the evaluation chain is four-step (genotype → phenotype → behaviour → fitness), and the behaviour depends on many external factors — the manual design of an autonomous and adaptive mobile robot is extremely difficult, so evolutionary approaches offer large potential benefits, including continuous and automated design, manufacture and deployment of robots of very different morphologies and control systems. Evolution is a slow learner, but it is inherently suited to parallelization, so more and more generations fit within practically acceptable timescales.

2. Morphological computation: the offloading debate

The contribution of the body to cognition and control in natural and artificial agents is increasingly described as "offloading computation from the brain to the body", where the body is said to perform morphological computation. Müller and Hoffmann (2017) investigate four characteristic cases and conclude that the offloading perspective is misleading: the contribution of body morphology to cognition and control is rarely computational, in any useful sense of the word.

The canonical statement is Pfeifer and Bongard (How the Body Shapes the Way We Think, 2007), with the sprawl robot exploiting the material properties of its legs for rapid locomotion, reducing the need for computation: "By 'morphological computation' we mean that certain processes are performed by the body that otherwise would have to be performed by the brain." The remarks lend themselves to two divergent interpretations:

  1. appropriate use of the body morphology leads to a reduction of the total amount of computation required to complete the task;
  2. the total amount of computation is constant, but a body takes over some of the computation originally performed by the brain and computes it in the morphology.
For the exam — the question to hold

In the example of walking, is this a computational task in the first place? Is the perspective of "offloading computation from the brain to the body" even feasible — a design decision one can take, choosing whether to compute centrally or in the body? The chapter's answer is a distinction, not a slogan: most "morphological computation" is morphology that facilitates control or perception; genuine morphological computation is extremely rare (reservoir computing is the clean case).

3. Four characteristic cases

  1. The passive dynamic walker — a minimal robot that can walk without any motors, control electronics, sensors or computer: two legs, a minimal torso, two arms; its ability to walk is exclusively due to the slope of the incline and the mechanical parameters of the walker (leg segment lengths, mass distribution, foot shape). The extreme example: if this is computing, what is not?
  2. Self-stabilizing machines, gecko feet, and the coffee balloon gripper — "active extensions" of the passive walker: the body is complemented by actuators, sensors and a simple control loop, yet the physical interaction of the body with the environment remains crucial. The gecko's feet allow walking on smooth vertical surfaces; the balloon gripper grips a complex object by simply enclosing it, the inner surface conforming passively — while a hard robotic hand must carefully compute how to grasp.
  3. The eye of the fly — perception: the distribution of light-sensitive cells is optimised for particular tasks, so a fly's eye facilitates perception — the morphology does part of the information processing.
  4. Physical reservoir computing — the case where the whole system, the body, literally seems to compute: a fixed, random dynamical substrate (physical or simulated) is driven by inputs, and only a readout is trained on its state.

4. Three classes: control, perception, and computation proper

Applying different notions of computation (abstract and physical) to the characteristic cases yields three classes:

ClassWhat it isCases
Morphology facilitating controlThe body's physics does work that the controller would otherwise have to do explicitly — but calling it "computation" adds nothingPassive dynamic walker; self-stabilizing machines; gecko feet; balloon gripper
Morphology facilitating perceptionThe sensory apparatus is shaped so that relevant information is pre-processed by the physicsFly eye; sensor placement and morphology in general
Morphological computation properThe body is actually used for computation — the substrate itself performs the information processingReservoir computing

The conclusion for design and analysis: focus not on computation, but on how morphology facilitates cognition and control — how it contributes to the overall orchestration of intelligent behaviour — and on how it can be specifically designed to do that, either by evolution (animals) or by engineers (soft robotics).

Widget — Reservoir computing: the substrate computes

A fixed random recurrent network (never trained) is driven by the input signal; only the linear readout is fitted to reproduce a target. This is morphological computation proper: the body does the computing, the controller only reads.

Widget — Which class?

Classify the devices into the three classes of the taxonomy. The widget corrects you with the reasoning of the paper.

5. Design implications: soft robotics and the body as a resource

The engineering reading is constructive: exploiting the body is a design strategy, not a workaround — the thread Chapter 1 promised to cash. A soft robot hand can grip a complex object simply by enclosing it; a robot built from stiff materials must apply complex control to adjust to uneven ground and will therefore be very slow (the sprawl-robot example). The growing field of soft robotics is the engineer's exploitation of the orchestration idea: machines that combine discrete rigid parts with continuous, soft materials, that can squeeze through holes, climb up walls, and flow around obstacles.

Controlling such devices is non-trivial — motion at one location propagates in unanticipated ways to other parts of the body — which is why evolution (Chapter 10) and automatic design (Chapter 11) are the natural design methods for them: Rieffel et al. successfully evolved locomotion for a soft robot such that it exploited rather than fought against the synergies within its body, and Bongard's tensegrity and soft robots appear precisely in the evolutionary robotics gallery.

Editor note

This is also where the course's frame-of-reference warning (Chapter 3) becomes constructive: if behaviour emerges from controller, body and environment, then changing the body is a way of changing the behaviour. Morphological computation is that idea taken seriously as engineering — and the online adaptation of the next sections is what you do when you cannot change the body but the body changes anyway.

6. Nanowire networks: a new substrate for control

The long-term goal of the Cesena group's research is the miniaturization of robots for missions in unknown environments. This limits the available energy, cost and size, making common computing systems unsuitable. One candidate: Nanowire Networks (NWNs) — novel neuromorphic devices with desirable characteristics: cheap to produce, small, and potentially integrating computing and memory. The drawback: their topological immutability — the network structure cannot be reprogrammed, so the device requires external interfaces and proper stimulation to induce the desired response, and each individual device is heterogeneous (no two are identical).

Robots controlled by such immutable, heterogeneous, network-based control systems cannot be "programmed" in the usual sense. What they can do is be adapted online: the robot acts in the environment, continuously adjusting its behaviour to improve its performance, maintaining a high performance during the whole life of the robot regardless of the changes that intervene during its activity. According to the dynamics of the environment, adaptation can work on different levels: switching between consolidated behaviours at the variance of external properties; continuously optimising some parameters to tune the behaviour; assembling pre-defined fundamental actions. The studies at hand use a further level: purely adaptive — the robot has no initial knowledge or ability and has to develop them from scratch; the sole information it has is an evaluation function representing the goal (biologically, comparable to the innate perception of pain or happiness that drives the very first behavioural adaptation in living beings).

7. Online adaptation: behaviour repaired in runtime

Online adaptation (OA) is the modification of the control software of a robot in runtime, with the aim of improving a performance measure. Differently from online learning techniques, the process of adaptation potentially never stops. Offline approaches — powerful and effective as they are — require caution: the produced controller might behave differently in the real world (the reality gap, Chapter 11), and, being immutable, it might not work properly in the face of changes in the robot's body (sensors and actuators undergo breakdowns or aging, potentially changing the resulting behaviour completely) and environmental conditions.

The key evaluation idea of the nanowire study: what matters is not only whether the robot finds an effective behaviour, but the life-long or cumulative performance — a robot able to perform a task efficiently in an instant but idle, useless or even damaging in the other moments is much less desirable than a mediocre robot that overall succeeds in its duty. The results: modifying the best found solution leads to constant improvement and to overall better cumulative performance, while a less constrained approach simplifies the exploration of different behaviours but reduces life-long performance — an experimental confirmation that the exploration/exploitation balance of Chapter 12 has a measurable, cumulative price.

ONLINE ADAPTATION — THE LOOP THAT NEVER STOPS ROBOT ACTS in the changing environment EVALUATION function scores the behaviour ADAPTIVE MECHANISM reconfigures the control in runtime never stops · maintains life-long performance · recovers from faults
Plate 16.2 — Online adaptation as a closed loop driven by an evaluation function. Contrast it with the offline loop of Chapter 11: there the design ends before deployment; here the design is the operation.

8. Fault recovery with Boolean-network robots

Baldini, Braccini and Roli (2025) ask the sharpest version of the question: can a robot recover from faults by itself? Two notions are distinguished: fault tolerance is the ability to maintain the desired behaviour despite damages (e.g. redundant sensors — the impact of a single damage is reduced, but performance is not necessarily recovered); fault recovery is the modification of some aspects of the robot or controller so as to recover the behaviour after the occurrence of damages — usually displaying a more abrupt decrease in effectiveness but, given some time, recovering performance to a level comparable to the original.

The scenario: minimal robots with constrained computational and power capabilities, controlled by Boolean Networks (BNs) — a model of Gene Regulatory Networks: a set of interconnected nodes that assume values in {0,1}; the value of a node depends on the values of k other "input" nodes through an update function; internal loops enable storing information like a memory (the system evolves according to its own past state, displaying a fading memory). The BN has 500 nodes; sensory inputs are binarized (values > 0.2 → 1), and the values of two BN nodes control the actuators. The adaptive mechanism is minimal: re-coupling — up to six inputs of the best-known coupling set are redirected to different BN nodes; if the new set attains better performance than the original set, it becomes the new best. The robot alternates exploration and exploitation phases, avoiding extended periods of underperformance while still allowing exploration. An embedded evaluation function drives the adaptation, enabling simple hardware implementation.

The results suggest that online adaptation is suitable for the automatic recovery of functions after the occurrence of damages — and, notably, adapting an existing controller to overcome a fault is faster than searching for a new controller from scratch: the same lesson as PLANEX (Chapter 14) and as macro-actions, now at the level of runtime adaptation on constrained hardware.

For the exam — the chapter-3 payoff

Chapter 3 flagged online evolution as the hard case (single robot, no population, no generations). Here is the workaround the field actually uses: keep the population inside the robot — a population of couplings/configurations explored in runtime, evaluated by an embedded function, with exploration/exploitation alternation. The four adaptations of Chapter 3 (evolutionary, learning, sensory, body) are all present in this one line of research: evolution (of couplings), learning (the evaluation-driven search), sensory and body adaptation (the damage is in the body; the re-coupling compensates it).

9. Reconfigurable organisms: from evolution to living machines

Kriegman, Blackiston, Levin and Bongard (2020) push the evolution of things to its biological limit. The claim: living systems are more robust, diverse, complex and supportive of human life than any technology yet created — and if living systems could be continuously and rapidly designed ab initio and deployed to serve novel functions, their innate ability to resist entropy might let them far surpass the useful lifetimes of static technologies. The pipeline is organized as a sequence of generators and filters:

  1. an evolutionary algorithm discovers different ways of combining the biological building blocks (cells) to realise the desired behaviour; a population of random designs is created, each design is simulated in a physics-based virtual environment and assigned a performance score; less-performant designs are deleted and overwritten by randomly modified copies of more-performant ones — the Chapter 10 cycle, applied to living tissue;
  2. transferable designs are then created using a cell-based construction toolkit to realize living systems with the predicted behaviours.

The resulting living systems are novel aggregates of cells that yield novel functions: above the cellular level, they bear little resemblance to existing organs or organisms. The same evolutionary algorithm is agnostic to the kind of artefact being designed — it can be reconfigured to design drugs, autonomous machines, metamaterials, or architecture. The possible applications named: safely delivering drugs inside the human body, environmental remediation, and broadening our understanding of the diverse forms and functions life may adopt.

Editor note

Read the pipeline once more: evolutionary search in simulation + physics-based evaluation + automated construction + deployment. That is the automatic off-line design loop of Chapter 11, with three substitutions — the "control software" is a body plan, the "simulation" is a physics-based environment of cells, and the "robot" is a living organism. The course's machinery did not change; the substrate did. That is the point of the evolution of things.

10. Closing the loop: the course's promises, cashed

Chapter 1 defined the robot through the sensory-motor loop and promised that exploiting the body is a design strategy — "Chapter 16 returns to it under the name of morphological computation". Here it is. The same chapter promised that behaviour must be measured statistically — Chapter 15 delivered the protocol. Chapter 2 promised Turing's child-machine — Chapter 10 delivered the evolutionary cycle. Chapter 3 promised that adaptation has four faces and that evolution is off-line by nature — this chapter showed the field working around the constraint (Symbrion in Chapter 9, online adaptation here). Chapter 4 promised that "no best controller" makes the experimental method compulsory — Chapter 15 made it precise. Chapter 9 promised that swarm robotics and automatic design are the course's two answers to robustness and flexibility — Chapter 11 delivered the second, and this chapter showed both converging on substrates nobody imagined in Chapter 1.

The course in one paragraph

Behaviour emerges from the interplay of controller, body and environment (Ch. 1). The frame-of-reference problem warns us not to attribute to the controller what the loop produces (Ch. 3). Architectures span the reactive–deliberative spectrum (Ch. 4): Braitenberg wirings (Ch. 5), subsumption (Ch. 6), motor schemas and fuzzy control (Ch. 7), behaviour trees (Ch. 8) — and when the design problem becomes too hard for a human, the course answers with self-organisation (Ch. 9), evolution (Ch. 10), automatic design (Ch. 11), learning (Ch. 12), deliberation (Ch. 13–14) — all validated by the experimental method (Ch. 15), and all pointing toward bodies, substrates and organisms that compute (Ch. 16). The thread is constant: autonomous and adaptive machines are designed by closing loops — through the world, through populations, through statistics — never by writing the behaviour directly.

Check your understanding

Describe the two major transitions of artificial evolution according to Eiben and Smith.

First transition (twentieth century): computer technology enabled artificial Darwinian processes in silico — evolutionary computation, the evolution of digital entities. Second transition (twenty-first century): developments in robotics, materials science and 3D printing enable the evolution of physical artefacts or machines — the evolution of things, evolution in materio. The Darwinian principles (variation, heredity, selection) are substrate-independent.

Define morphological computation and the two interpretations of the offloading idea.

Morphological computation is the idea that certain processes are performed by the body that otherwise would have to be performed by the brain (Pfeifer & Bongard). Two interpretations: (1) appropriate morphology reduces the total amount of computation required; (2) the total computation is constant but the body takes over part of it. Müller and Hoffmann argue the offloading perspective is misleading: the body's contribution is rarely computational in any useful sense.

List the four characteristic cases and the three classes of the taxonomy.

Cases: the passive dynamic walker (no motors, no sensors, no computer); self-stabilizing machines, gecko feet and the coffee balloon gripper; the eye of the fly; physical reservoir computing. Classes: morphology facilitating control; morphology facilitating perception; morphological computation proper (rare — reservoir computing is the clean case).

Explain why the coffee balloon gripper is a case of morphology facilitating control.

The gripper grips a complex object by simply enclosing it: the inner surface of the hand passively conforms to the object. A hard robotic hand must carefully compute how to grasp. The body does work the controller would otherwise have to do explicitly — the total computation is reduced because the morphology contributes to the orchestration of the behaviour, but nothing is "computed" in the substrate.

What is reservoir computing, and why is it the case of morphological computation proper?

A fixed, random dynamical substrate (physical or simulated) is driven by inputs, and only a readout is trained on its state — the substrate itself performs the information processing. This is the case where the whole system literally seems to compute, hence the cleanest example of morphological computation proper.

Why are nanowire networks an interesting control substrate, and what is their drawback?

Nanowire networks are cheap to produce, small, and potentially integrate computing and memory — suitable for miniaturized robots with constrained energy, cost and size. The drawback: their topological immutability — they cannot be reprogrammed, so they require external interfaces and proper stimulation to induce the desired response, and each device is heterogeneous.

Define online adaptation and explain the purely adaptive level.

Online adaptation is the modification of the control software of a robot in runtime, with the aim of improving a performance measure; unlike online learning, the process potentially never stops. Levels: switching between consolidated behaviours; continuously optimising parameters; assembling pre-defined actions. The purely adaptive level: the robot has no initial knowledge or ability and develops them from scratch, with the sole information coming from an evaluation function representing the goal.

Why does life-long performance matter in online adaptation, and what did the nanowire study find?

Because a robot that performs efficiently in an instant but is idle or harmful the rest of the time is worse than a mediocre robot that overall succeeds: the whole life of the agent matters. The study found that modifying the best found solution leads to constant improvement and better cumulative performance, while a less constrained approach simplifies exploration but reduces life-long performance.

Distinguish fault tolerance from fault recovery, and describe the Boolean-network adaptation mechanism.

Fault tolerance maintains the desired behaviour despite damages (e.g. redundant sensors — impact reduced, performance not necessarily recovered). Fault recovery modifies some aspect of the robot or controller to recover the behaviour after damage — a more abrupt decrease but eventual recovery. The BN mechanism: a 500-node Boolean network (values in {0,1}, nodes updated by Boolean functions of k inputs, internal loops as fading memory) receives binarized sensor inputs; the adaptive mechanism re-couples up to six inputs of the best-known coupling set to different BN nodes, keeping the new set if it performs better; exploration and exploitation phases alternate, driven by an embedded evaluation function. Adapting an existing controller to overcome a fault is faster than searching from scratch.

Describe the reconfigurable-organisms pipeline of Kriegman et al. (2020).

A linear pipeline of generators and filters: an evolutionary algorithm designs diverse candidate lifeforms in silico by combining biological building blocks (cells), each design simulated in a physics-based virtual environment and scored; less-performant designs are deleted and overwritten by mutated copies of more-performant ones; transferable designs are then manufactured with a cell-based construction toolkit to realize living systems with the predicted behaviours. The resulting organisms are novel aggregates of cells bearing little resemblance to existing organs — with potential applications in targeted drug delivery, environmental remediation, and the study of life's possible forms.

How does the reconfigurable-organisms pipeline echo the automatic off-line design loop?

It is the same loop with three substitutions: the "control software" is a body plan; the "simulation" is a physics-based environment of cells; the "robot" is a living organism. Evolutionary search in simulation + physics-based evaluation + automated construction + deployment — the machinery of Chapter 11 did not change, the substrate did. That is the point of the evolution of things.

Summarise the course in one paragraph, per the closing section.

Behaviour emerges from the interplay of controller, body and environment; the frame-of-reference problem warns against attributing to the controller what the loop produces; architectures span the reactive–deliberative spectrum; when design becomes too hard for a human, the course answers with self-organisation, evolution, automatic design, learning and deliberation — all validated by the experimental method and all pointing toward bodies, substrates and organisms that compute. The thread: autonomous and adaptive machines are designed by closing loops — through the world, through populations, through statistics — never by writing the behaviour directly.