Evolving virtual creatures

TL;DR: In 1994 Karl Sims evolved swimming, walking, and light-chasing creatures inside a physics simulator, optimizing body plan and neural controller together from a recursive graph genotype — and got designs no engineer would have drawn. The idea then starved for thirty years because physics simulation and evolutionary search each needed orders of magnitude more compute than existed, and there was no way to build the winners. GPU physics now supplies roughly the missing 10⁶×, differentiable simulators and 3D printing close the loop to hardware, and yet the robotics industry has converged on one humanoid form factor without ever seriously running the search. Morphology remains the least exploited optimization lever in robotics, and this paper is its founding document.

The idea as Sims had it

The paper makes one claim that still sounds radical: the body is part of the policy. If you fix a robot's morphology and then optimize a controller, you are solving a constrained version of the real problem. Sims instead defined a genotype that specifies both at once and let evolution search the joint space.

The genotype is a directed graph. Each node describes a rigid body part — dimensions, joint type connecting it to its parent (rigid, revolute, twist, universal, spherical…), and joint limits. Each edge describes how a child part attaches: position, orientation, scale, reflection. Crucially, the graph can contain cycles and self-loops with a recursion limit, so a single node with a self-edge unrolls into a chain of segments, and a reflected edge yields bilateral symmetry for free. A compact genotype of a handful of nodes expands into a phenotype of dozens of parts. This is a developmental encoding: it builds in reuse, symmetry, and segmentation — the things bodies actually have — rather than hoping mutation rediscovers them limb by limb.

The controller lives inside the same graph. Each node carries a local neural circuit — not just sums-and-sigmoids but a zoo of function units (oscillators, integrators, min/max, interpolation, memory) — wired to local sensors (joint angle, contact, photosensors) and effectors that torque the joints. When a body node is replicated by recursion, its neural circuitry replicates with it, so a segmented tail automatically gets a segmented, coupled controller. A central non-replicated circuit coordinates. Brain topology mirrors body topology by construction.

Evolution is then almost embarrassingly plain: populations of roughly 300, on the order of 100 generations, mutation on node/edge parameters and graph topology, crossover by grafting subgraphs, fitness measured by simulating each creature in rigid-body physics with collisions and (for swimmers) fluid drag. Fitness functions were distance swum, distance walked, jump height, or ability to follow a light source. A companion paper the same year (Evolved 3D Morphology and Behavior by Competition, Artificial Life IV) added coevolutionary arms races over possession of a cube — arguably the first open-ended-ish competitive embodied evolution.

The results were the point. Evolution produced paddling turtles, undulating snakes, tumbling and shuffling walkers, and creatures exploiting solutions no one specified. It also produced the now-canonical cautionary tale: creatures that got "fitness" by exploiting integrator bugs and energy-nonconserving contacts — the original reward hacking, a decade before anyone called it that.

Why it could not work in 1994

Put numbers on the wall Sims hit.

Simulation throughput. He ran on a Connection Machine CM-5 — a national-lab-class parallel supercomputer — using, as I recall the paper, a few tens of processors, with peak throughput in the low gigaflops. Each fitness evaluation required simulating articulated rigid-body dynamics with collision detection for several seconds of virtual time. A full run (order 30,000 evaluations) took hours on the CM-5; a research program of many runs took days-to-weeks per question. Today a single H100-class GPU delivers tens of teraflops of general FP32 — roughly four orders of magnitude over the whole CM-5 partition — and, more importantly, GPU-native simulators (Isaac Gym/Isaac Lab, Brax, MuJoCo MJX) batch thousands of environments per device, sustaining roughly 10⁵–10⁶ environment-steps per second. The end-to-end throughput gap between Sims' setup and one modern GPU is plausibly a factor of ~10⁵–10⁶, exactly the "orders of magnitude" the method starved for.

Simulation fidelity. Contact models of the era were penalty-based and leaky; evolution, being an adversarial optimizer against your simulator, found every leak. Sims had to hand-patch exploits (creatures vibrating to harvest integrator energy). Without robust contact solvers and domain randomization, evolved gaits were artifacts of one simulator's bugs.

Search budget. 300 individuals × 100 generations is a laughably small budget by modern standards — modern RL burns 10⁹ steps on a single fixed morphology. Sims was searching a joint body–brain space with fewer total evaluations than one PPO run uses per hour.

No path to hardware. Even if you evolved a great walker, you could not build it: no cheap 3D printing, no modular actuators, no sim-to-real methodology. The output was a movie for SIGGRAPH, and that framing — computer graphics, not robotics — is partly why the idea stalled.

What changed

Four things, each independently transformative:

1. GPU physics. Batched, GPU-resident rigid-body simulation gives you the 10⁵–10⁶× throughput multiplier. Population-based search over morphologies is now an overnight job, not a supercomputer allocation.
2. Differentiable simulation. Brax, Warp, DiffTaichi and successors expose gradients of trajectories with respect to design parameters — link lengths, densities, joint placements. Sims had to treat morphology as a black box; we can locally hill-climb continuous morphology while reserving evolution for discrete topology changes, which is where graph mutation actually earns its keep.
3. Amortized control. Sims evolved each creature's controller from scratch, so a promising body mutation looked bad until its brain caught up — the central pathology of body–brain coevolution (Cheney, Bongard and colleagues later named and dissected this: controllers become entrenched in a morphology, punishing body innovation). Morphology-conditioned policies — transformers over the body graph, one policy for a whole family of bodies (MetaMorph and kin) — mean a mutated body arrives with a competent controller on day one. This attacks the algorithmic bottleneck, not just the compute one.
4. Fabrication. Cheap multi-material printing plus standardized actuator modules mean an evolved design can exist. Lipson and Pollack's Golem project (2000) printed evolved robots at toy scale; Kriegman, Bongard and colleagues' xenobots (2020) evolved designs realized in living cells. The loop closes.

What a serious 2026 revival looks like

Keep from Sims: the graph-grammar genotype. It is still better than most modern encodings — voxel grids don't compose, and direct parameter vectors can't express "add one more segment, with its controller." Modern grammar-based systems (MIT's RoboGrammar, 2020) rediscovered essentially this representation.

Replace: the inner-loop controller evolution with a bilevel scheme. Outer loop: population search over body graphs (mutation, subgraph crossover, plus an LLM as a semantically informed mutation operator proposing graph edits — "give it a counterweight tail" — which early work on LLM-guided evolutionary search suggests is a real accelerant, though evidence in morphology design is thin). Inner loop: a shared morphology-conditioned transformer policy fine-tuned briefly per candidate with PPO in ~10⁴ parallel Isaac Lab environments, plus differentiable-simulation gradients on continuous design parameters. Fitness evaluated under domain randomization and multiple terrains so evolution can't overfit one contact model — the modern answer to Sims' energy-leak exploiters.

Close the loop: constrain the grammar to a catalog of printable links and off-the-shelf actuators (quasi-direct-drive modules), auto-generate CAD from the winning graph, print and assemble the top-k designs monthly, and feed real-world telemetry back as a fitness correction. Total cost is a few GPUs and a print farm — a startup-sized bet, not a moonshot.

The pointed question such a program answers: is the humanoid form factor, on which billions are currently being bet, actually near-optimal for warehouse and household tasks — or merely anthropomorphically convenient? Nobody has run the search at scale. Sims' paper says the search finds things designers don't.

Descendants, vindication, and what's still open

The lineage is real but thin. Lipson & Pollack (2000) put evolved bodies in hardware. Cheney et al. (2013) evolved soft voxel creatures with CPPN encodings. RoboGrammar (2020) did grammar-based morphology search with learned heuristics. Gupta et al.'s DERL (2021) ran large-scale body–brain evolution with RL inner loops and showed morphology shapes learnability itself. Transform2Act and EvoGym made the co-design problem RL-native and benchmarkable. Xenobots took the idea into biology. So: partially vindicated, repeatedly, at small scale.

What is not settled: (a) whether co-design search, given modern compute, escapes the entrenchment local optima or just finds them faster — reported results still show morphology search converging early; (b) whether evolved morphologies survive sim-to-real for contact-rich dynamic gaits, which nobody has demonstrated convincingly beyond quasi-static or soft-body cases; (c) open-endedness — Sims' competitive coevolution hinted at arms races that no one has scaled, and it may be the most interesting thread left dangling.

Where to read it

The paper is at doi.org/10.1145/192161.192167 (bibliographic details verified; my recollections of exact run times and processor counts are from memory of the text and should be treated as approximate). Watch the original video — it carries half the argument. Read alongside: Sims' companion Artificial Life IV competition paper (1994); Lipson & Pollack, "Automatic design and manufacture of robotic lifeforms" (Nature, 2000); Cheney et al. on morphological entrenchment in co-optimization (~2016–2018); RoboGrammar (SIGGRAPH Asia 2020); and Gupta et al., "Embodied intelligence via learning and evolution" (2021), which is the closest thing yet to the revival sketched above — still run at a scale Sims would recognize, on hardware he wouldn't believe.