Evolving Machines: When Bodies Compute and Morphology Is the Algorithm

## The Assumption We Inherited

Classical robotics inherited an assumption from the computing paradigm of the mid-twentieth century: the body is a passive executor of instructions issued by the brain. Intelligence lives in the controller; the body is merely the effector. This assumption produced robots of a particular kind---rigid, precisely actuated, heavily sensored, computationally expensive to control. It also produced a deep and recurring frustration, because the natural systems that robots were meant to emulate---animals, insects, plants in motion---do not appear to operate this way at all.

The alternative view, articulated with theoretical clarity by Pfeifer and Bongard in their 2006 synthesis *How the Body Shapes the Way We Think*, is that the body is not peripheral to cognition but constitutive of it (Pfeifer & Bongard, 2006). Morphology is a computational resource. The geometry of a limb, the compliance of a joint, the distribution of mass---these are not constraints to be worked around by a clever controller but features that perform computation in their own right. The central thesis of what has come to be called morphological computation is simple and radical: good design offloads cognitive work onto physical structure.

A passive dynamic walker---a device with no actuators, no sensors, no controller---descends a slight incline with a gait that is visually indistinguishable from human walking. The computation required to produce that gait resides entirely in the geometry and mass distribution of the mechanical system. No algorithm generates it. The body is the algorithm.

### Co-Evolution of Body and Brain

If morphology performs computation, then the form of a body and the structure of its controller cannot be designed independently. They co-determine one another. This intuition acquired experimental grounding in a landmark demonstration by Lipson and Pollack, who evolved both the body plans and neural controllers of three-dimensional locomoting robots from scratch, using only the raw materials of physics and selection (Lipson & Pollack, 2000, *Nature*, 406:974--978).

The Golem project, as it was called, used evolutionary algorithms to search simultaneously through morphological space and control space. Starting from random assemblages of bars, joints, and neurons, the system discovered coherent locomoting bodies---creatures that had never been designed by any engineer and bore no deliberate resemblance to biological precedents. Some undulated. Some rolled. Some walked in gaits that, on close inspection, revealed mechanical tricks: a particular limb geometry that passively stores and returns elastic energy, reducing the load on the neural controller. The body had learned to help.

The significance of this result extends beyond its novelty. Lipson and Pollack showed that evolution, given enough time and the right fitness criterion, navigates toward solutions in joint morphology-control space that engineering intuition reliably misses. The space of possible bodies is too large to explore by human design. Evolution is not a heuristic for finding approximate solutions; for problems of sufficient complexity, it is the only known method that reliably finds solutions at all.

### Compliance as a Feature

The shift from rigid to soft robotics represents, in retrospect, a reassertion of the morphological computation principle at a material level. Rigid robots require precise, computationally expensive control because they have no mechanical tolerance for error---a deviation of a few millimetres in a rigid linkage propagates as a positional error through the entire kinematic chain. Soft, compliant bodies absorb such deviations passively. The material itself performs error correction.

This is not a marginal engineering convenience. It is a fundamental repartitioning of the computational load between controller and body. A soft robotic gripper with silicone fingers can grasp objects of arbitrary geometry without any explicit sensing of that geometry: the fingers deform to fit the surface. The computation required to determine the correct gripping configuration---which would demand sophisticated perception, modelling, and trajectory planning in a rigid system---is performed mechanically, in microseconds, by the passive deformation of the material. The controller becomes simpler precisely because the body has become more capable.

Soft robotics has further revealed that locomotion strategies inaccessible to rigid systems become natural in compliant materials. Pneumatically actuated octopus-inspired arms, continuum robots derived from the mathematics of flexible slender bodies, and crawling robots whose gait emerges from coupled oscillations in their elastic structure all demonstrate that movement strategies borrowed from the animal kingdom become mechanically natural when you abandon the assumption of rigidity (Rus & Tolley, 2015, *Nature*, 521:467--475).

### Living Machines

The most extreme expression of the morphological computation principle is a robot whose body is not engineered material at all but biological tissue. In 2020, Kriegman and colleagues reported the creation of xenobots: autonomous agents assembled from dissociated frog embryo cells and arranged according to computer-designed blueprints that specify which cells should become muscle and which should become scaffold (Kriegman et al., 2020, *Proceedings of the National Academy of Sciences*, 117:1853--1859).

The design process was explicitly evolutionary. A population of candidate body plans was evaluated in simulation for locomotor performance, with the highest-performing designs selected for iteration. The evolutionary algorithm explored a combinatorial space of cell arrangements that no human designer would have navigated---not because the individual arrangements are difficult to reason about, but because the space is simply too large for intuitive search. The resulting xenobot designs are, in their geometry, genuinely non-intuitive. They look like compressed lumps, not like any robot or any animal. And yet they move with reliable directionality, traverse confined microfluidic environments, and can be made to transport small payloads.

What the xenobots represent is a conceptual boundary dissolution. They are not robots in the conventional sense---they have no engineered actuators, no electronic sensors, no circuit board. They are not organisms in the conventional sense---they are assemblies designed by an evolutionary algorithm, not products of selection across generations. They occupy a category that existing terminology does not cleanly describe: designed biological machines whose computational architecture is entirely embodied, whose locomotion strategy is entirely morphological, and whose material substrate is living tissue with its own metabolic agenda.

A subsequent study demonstrated that xenobots, when placed in proximity to loose frog cells, spontaneously assembled those cells into new xenobots---a form of kinematic self-replication that was itself discovered by the evolutionary algorithm rather than deliberately designed (Kriegman et al., 2021, *Proceedings of the National Academy of Sciences*, 118:e2112672118). The evolutionary search had found a reproductive strategy without being asked to.

### Recovering from Damage Without a Map

A persistent limitation of conventional robotics is brittleness in the face of unanticipated damage or environmental change. A legged robot with a broken actuator, operating from a controller designed for its intact body, will fail to locomote---the controller has no model for the new mechanical reality. This failure mode is foreign to biology. Animals with damaged limbs compensate spontaneously, adopting altered gaits that preserve locomotor function. The compensation does not require explicit damage detection and model revision; it emerges from the interaction of a distributed nervous system with a body that provides ongoing mechanical feedback.

Cully and colleagues addressed this challenge directly, introducing a quality-diversity (QD) algorithm called MAP-Elites to generate a behavioural repertoire that an injured robot could search rapidly at runtime (Cully et al., 2015, *Nature*, 521:503--507). The approach has two phases. Before deployment, MAP-Elites runs offline and generates a large, structured archive of locomotion strategies, each characterised by its performance and by descriptors of the robot's behaviour---for instance, the fraction of time each limb is in contact with the ground. This archive is not a single optimal policy but a map of the space of possible behaviours, covering many diverse and high-quality solutions simultaneously.

At runtime, when the robot detects a drop in locomotor performance consistent with damage, it searches the pre-generated behavioural archive using an adaptive sampling strategy that quickly identifies which archived behaviours transfer well to the robot's current state. In experiments, a six-legged robot with multiple actuator failures recovered functional locomotion within minutes by navigating its behavioural repertoire---not by diagnosing the damage or re-optimising from scratch, but by quickly discovering which of the pre-computed strategies happened to work given the altered morphology.

The MAP-Elites result matters beyond the specific robotics application because it is a demonstration that evolution-inspired algorithms can produce knowledge structures---behavioural archives---rather than just solutions. The archive is an externally stored map of solution space that an agent can consult adaptively. This is a different relationship between an evolutionary algorithm and an agent than the classical one in which evolution produces a fixed policy. Here, evolution produces a flexible resource.

### What Engineering Intuition Cannot Reach

The accumulated evidence from evolutionary robotics converges on a lesson that is uncomfortable for engineering culture but hard to evade: evolution explores design spaces that human intuition cannot reach, not because human designers are insufficiently clever, but because the spaces are genuinely too large and too counterintuitive to navigate by inspection.

This is not a failure of engineering. It is a structural property of high-dimensional design spaces. When the relevant dimensions include morphology, material compliance, neural architecture, and developmental trajectory simultaneously, the interactions between these dimensions produce fitness landscapes with structures that cannot be usefully approximated by local search or by decomposition into subproblems. Evolution is a global search procedure that does not decompose the problem---it evaluates wholes, and it is indifferent to which dimension of the design space a particular improvement came from.

The practical implication is that designing machines for tasks that require adaptive behaviour in unstructured environments---the tasks for which biological organisms are optimised---will increasingly require evolutionary methods. This is not an admission that we cannot understand the solutions evolution finds; it is a recognition that understanding solutions post hoc is different from finding them in the first place. The xenobot body plans are, after the fact, interpretable. No human would have generated them a priori.

The longer arc of evolutionary robotics is a progressive erosion of the boundary between the biological and the engineered. When the bodies of machines are grown from living cells, when their controllers are evolved rather than programmed, when their behavioural repertoires are pre-computed diversity landscapes rather than fixed policies, the distinction between robot and organism becomes a matter of substrate and origin rather than of principle. The organisms that evolution produced over three billion years were, from the beginning, machines whose bodies compute. We are slowly learning to build them deliberately.

---

**Key References**

- Pfeifer, R., & Bongard, J. (2006). *How the Body Shapes the Way We Think: A New View of Intelligence*. MIT Press. - Lipson, H., & Pollack, J.B. (2000). Automatic design and manufacture of robotic lifeforms. *Nature*, 406, 974--978. https://doi.org/10.1038/35023115 - Rus, D., & Tolley, M.T. (2015). Design, fabrication and control of soft robots. *Nature*, 521, 467--475. https://doi.org/10.1038/nature14543 - Kriegman, S., Blackiston, D., Levin, M., & Bongard, J. (2020). A scalable pipeline for designing reconfigurable organisms. *Proceedings of the National Academy of Sciences*, 117, 1853--1859. https://doi.org/10.1073/pnas.1910837117 - Kriegman, S., Blackiston, D., Levin, M., & Bongard, J. (2021). Kinematic self-replication in reconfigurable organisms. *Proceedings of the National Academy of Sciences*, 118, e2112672118. https://doi.org/10.1073/pnas.2112672118 - Cully, A., Clune, J., Tarapore, D., & Mouret, J.-B. (2015). Robots that can adapt like animals. *Nature*, 521, 503--507. https://doi.org/10.1038/nature14422

Back to blog