challenge accepted

Why virtual creatures matter

T
he first virtual creatures were
evolved by Karl Sims in the early
1990s1: morphology and control of
autonomous agents were co-optimized so
that they ran, jumped, swam, performed
phototaxis and fought head-to-head for
resources in a virtual world that followed
(more or less) the laws of classical An evolved, galloping creature. Adapted from ref. 4, ACM.
mechanics. The creatures were relatively
simple, composed of just a handful of
jointed, rigid components, but their much greater morphological possibility: they in addition to actuating, can choose at every
behaviour was surprisingly rich and lifelike2. can change shape and material properties, as timestep to either attach to their nearest
Karl Sims’s experiments started with well as their control policies. neighbour (forming an aggregate, symbiotic
an objective, such as locomotion, and a By instantiating soft machines in a virtual machine with a shared reward function) or
population of randomly assembled creatures. world, it is possible to rapidly and cheaply detach, reconfigure and test a new design.
Although it was unlikely that any randomly evaluate many candidate designs in parallel. Thus morphology is controlled by the same
assembled creature would fully satisfy the This makes exploration of the design space policy that coordinates behaviour.
objective, by replacing the worst-performing much more efficient than if each proposed What comes next for virtual creatures
designs with slightly and randomly modified design needed to be built in reality, and research may be to distil and combine the
copies of the better ones, the population permits us to use optimization algorithms to most useful properties of evolutionary and
made incremental progress, generation by automatically discover and refine a robot’s gradient-based algorithms, and discard the
generation. It was the survival of the fittest; shape, material properties and controllers rest. However, this will require some finesse.
or, in the case of locomotion: the fastest. contemporaneously, thus leading to entirely More powerful optimization, without
In 2000, Lipson and Pollack3 transferred novel solutions outside the purview of recourse to testing in the real world, might
similarly evolved creatures from simulation human intuition5. result in overfit virtual creatures forever
to reality with a technology just then However, the goal of designing virtual trapped inside the computer.
emerging: 3D printing. Robot designs were creatures is not necessarily to design the best But even completely virtual life can hold
rapidly and safely prototyped as virtual soft robot. The goal can also be to engage real value. And this is precisely where virtual
creatures, discarding the truly awful or our imagination and creativity — to expand creatures depart from the current paradigm
dangerous designs before testing them in our conception of life and intelligence, and of most machine learning research: virtual
reality. The result were ‘robotic lifeforms’ that how these things might materialize. creatures are not constrained by their
were designed, optimized and built, end-to- Indeed, Sims, who was primarily interested applicability. They are interesting objects of
end, with almost no human intervention. in computer graphics, viewed his creatures scientific investigation in their own right.
Since then, despite many attempts to as artistic expressions, but others (the author More than that, they have the potential to be
improve on this seminal work, and despite of this essay included) saw his work as an as beautiful and complex as life itself. ❐
vast increases in computational power, exciting new paradigm in the science of AI.
there has been relatively little progress Since its inception, the Virtual Creatures Sam Kriegman
scaling the complexity, competence and Competition has been affiliated with the University of Vermont, Burlington, VT, USA.
usefulness of virtual creatures and their Genetic and Evolutionary Computation e-mail: [email protected]
3D-printed equivalents. Conference (GECCO), a machine learning
This stagnancy was ended in 2013 with conference inspired by evolution and ecology Published online: 9 October 2019
an important advance by Cheney et al.4 — rather than learning and neuroscience. This https://doi.org/10.1038/s42256-019-0102-8
winners of the inaugural Virtual Creatures is because morphological optimization
Competition — which introduced the proper — changing the number and References
1. Sims, K. Artif. Life 1, 353–372 (1994).
evolution of soft robots composed of placement of mechanical degrees of freedom,
2. Sims, K. Evolved Virtual Creatures (1994); https://archive.org/
thousands of elastic voxels (pictured). This not merely tuning the parameters of a details/sims_evolved_virtual_creatures_1994
work became highly influential, and for predefined structure — has thus far only been 3. Lipson, H. & Pollack, J. B. Nature 406, 974–978 (2000).
good reason. demonstrated through evolutionary design. 4. Cheney, N., MacCurdy, R., Clune, J. & Lipson, H. in Proc. Genetic
and Evolutionary Computation Conference 167–174 (ACM, 2013).
Soft robots provide a unique opportunity The winners of this year’s competition, 5. Thompson, A. in Proc. International Conference on Evolvable
for virtual creatures research: soft structures Deepak Pathak and colleagues6, have shattered Systems (eds Higuchi, T., Iwata, M. & Liu, W.) 390–405
can act as shock absorbers, energy stores this assumption by demonstrating how the (Springer, 1996).
6. Pathak D. et al. Preprint at https://arxiv.org/abs/1902.05546 (2019).
and even power amplifiers (catapults). morphology and controller of a machine
However, they are extremely non-intuitive can be co-optimized using a gradient-based
Acknowledgements
to design and control due to low mechanical learning algorithm, without evolution. Instead Thanks to S. Beaulieu and J. Bongard for helpful suggestions.
impedance (force applied to one part of the of evolving a monolithic morphology,
robot can propagate in unanticipated ways Pathak et al. optimize a swarm of elemental Competing interests
throughout its body), and because they have agents — autonomous ‘limbs’ — that, The author declares no competing interests.

492 Nature Machine Intelligence | VOL 1 | OCTOBER 2019 | 492 | www.nature.com/natmachintell