Nature Got There First: What 3.8 Billion Years of Biomimicry Can Teach Physical AI

Humanoid robots competed in gymnastics at the World Humanoid Robot Games in Beijing last month: flips, balance, the kind of full-body routine that would have been unthinkable for a machine a few years ago.^[9] Some of that came from the same underlying trick behind a piece of research that came out around the same time: a robot that taught itself to cartwheel, not by being programmed to cartwheel, but by watching 2.5 hours of human motion data, learning the underlying shape of how bodies move, and recombining what it learned into moves nobody explicitly trained it to do. It went from walking, into an aerial cartwheel, into a spin kick, and back to walking, hitting rotational speeds comparable to a trained human athlete.^[1] None of this is really a robotics story on its own. It's a biomimicry story, and it's a good excuse to talk about something I think gets undersold in most of the Physical AI conversation: nature isn't a mood board for robot design. It's the original research programme, and it's been running for 3.8 billion years.

1. Model, Measure, Mentor

Janine Benyus, who named this field in her 1997 book, put it more precisely than most engineers manage: nature as model, nature as measure, nature as mentor.^[2] Model, because organisms have already solved the design problem you're working on. Measure, because whatever nature settled on after billions of years of selection pressure is a really useful benchmark for whether your solution is any good. Mentor, because the point isn't to copy a shape, it's to learn the underlying principle well enough to apply it somewhere the original organism never had to.

That third one is where most biomimicry attempts go wrong, so let me flag it before I get into the good examples. A chair leg that looks like a bird bone is biomorphic, not biomimetic. It's borrowing the aesthetic, not the engineering. Real biomimicry means understanding why a gecko's foot grips glass, why a termite mound stays a constant temperature without any active cooling, why a school of fish can turn as one unit with no leader giving the order, and then applying that underlying principle to a problem nature never had to solve, like a robot that has to climb a wet pipe or a data centre that has to stay cool without burning through a power budget.

2. The Body: Movement Nature Already Solved

Physical AI's biggest unsolved problem, in terms of the Act pillar of the Sense - Reason - Act framework I've used throughout this year's articles, is still locomotion and manipulation in unstructured environments. Wheels are a human invention nature never really needed. Legs, grip, and soft-bodied dexterity are nature's answer, refined across an enormous range of terrain and predator pressure, and Physical AI is increasingly borrowing the answer rather than reinventing it.

The cartwheeling robot I opened with, controlled by BeyondMimic, is one end of this: it learns human agility directly from motion capture, then uses a two-stage process: reinforcement learning to track the movement, then a latent diffusion model to recombine and adapt it. That way it can generate a new skill on the fly rather than needing to be separately trained for every single manoeuvre.^[1] In a study of 77 people watching its walking and running gait against a standard robot controller, people picked BeyondMimic's version as more humanlike and natural 71% of the time.^[1] That's not a coincidence. It's the whole point. Human movement, refined over millions of years for efficiency and balance, is a much better answer than most hand-coded gait algorithms.

Further down the food chain, quite literally, researchers at Carnegie Mellon combined two entirely different biological mechanisms into one soft robot: a sea star's crawling gait for locomotion, and a gecko's adhesive foot pads for grip, giving a soft, amphibious robot the ability to climb wet and dry surfaces up to a 51-degree slope.^[3] Neither mechanism on its own was the answer. The combination was. That's a really useful reminder for anyone doing biomimicry in an engineering context: nature rarely hands you one clean solution, it hands you a library of partial solutions, and the engineering work is in the recombination.

3. The Beijing Proof Point

If you want to see biomimicry's speed of progress and its actual limits in the same five days, you didn't need to look further than Beijing last month. The second World Humanoid Robot Games ran from 22nd to 26th August 2026, with more than 2,000 humanoid robots from 16 countries competing in Olympics-style events: sprinting, football, boxing, dance, and a full slate of "work" events like firefighting drills, library sorting, and dexterous manipulation.^[9]

The headline number is startling. A robot called Tiangong Ultra, built by Beijing's X-Humanoid, ran the 100 metres in 8.64 seconds, comfortably faster than Usain Bolt's 9.58-second human world record, after records fell on consecutive days across the week: 9.39, then 9.34, then 9.32, then 8.86, then 8.64.^[10] Twelve months earlier, at the first Games in 2025, the winning 100-metre time was 21.50 seconds. The record was cut nearly in half in a single year.^[11] The same robot broke the human 400-metre world record too, running 38.15 seconds.^[10]

And then, repeatedly, the winning robots crashed straight into the padded stopping barrier at the end of the track. Several caught fire. Officials ran out with extinguishers while robots were carried off on stretchers.^[12] One piece of coverage landed on the line that I think captures the whole point of this article better than I could: the robots beat Usain Bolt, then met their match: picking up beans with tweezers.^[13] Sprinting in a straight line down a marked lane is a narrow biomechanical problem, and Physical AI has essentially solved it. Stopping gracefully, recovering balance after an unexpected stumble, and manipulating a small irregular object with a soft grip are all things a three-year-old does without thinking about it, and they remain some of the hardest unsolved problems in the entire field.

4. The Brain: Computing the Way Nervous Systems Actually Compute

The part of biomimicry that gets the least public attention is the one I'd bet on mattering most for where Physical AI is heading: neuromorphic computing. Every GPU running a modern AI model is built on the von Neumann architecture: memory in one place, processing in another, data constantly shuttling between them, burning power the entire time. A biological brain doesn't work that way at all. Memory and processing are the same physical location, information moves as brief electrical spikes rather than continuous values, and the whole system runs on roughly 20 watts, about what a dim light bulb draws.^[4]

Neuromorphic chips copy that architecture directly rather than just borrowing the name. Intel's Loihi 2 packs over a million artificial neurons onto a chip that only activates the circuits relevant to whatever input it's currently processing, the same way a biological neuron only fires when it actually needs to.^[4] In 2026 these chips are moving out of research labs and into exactly the applications this series has spent the year on: autonomous robotics that need sub-millisecond reaction times and all-day battery life, edge AI that can't rely on a constant cloud connection, sensors that need to learn and adapt on the device itself rather than shipping data back to a data centre to be retrained.^[4] This is, in terms of the data sovereignty argument I made earlier in this series, a really important development: an architecture that's efficient enough to reason at the edge, on-device, is also an architecture that keeps sovereign data sovereign, because it never has to leave the device to be useful.

5. The Colony: Intelligence That Doesn't Live in One Place

Ant colonies solve logistics problems that would take a supercomputer a long time to brute-force, and they do it with no individual ant understanding the overall plan. Each ant follows a simple local rule: reinforce the pheromone trail that led to food, let unused trails fade. And the colony as a whole finds close to the shortest route between nest and food source. Ant colony optimisation, the algorithm directly modelled on that behaviour, is now a standard tool for path planning in mobile robotics, still being actively refined in 2026 to handle the kind of complex, dynamic environments a real warehouse or disaster site throws up.^[5]

The wider field this sits inside, swarm intelligence, is explicitly built on the same pattern found in bird flocks, fish schools, and insect colonies: no central controller, just simple local rules producing coordinated, resilient group behaviour.^[6] That's a very different design philosophy to most enterprise AI deployment, which still tends to centralise: one model, one point of control, one point of failure. A swarm of drones inspecting a wind farm, or a fleet of warehouse robots rerouting around a blocked aisle without checking in with a central controller, is applying millions of years of insect and fish evolution to a coordination problem that's still really hard for centralised software architectures to handle gracefully.

6. The Real Argument: World Models Are a Biomimicry Project Too

Here's the idea I want to leave you with, because I think it reframes the whole series rather than just adding a chapter to it. A world model, in the sense I've used the term across this year's articles, is a system that builds an internal representation of its environment and predicts how that environment changes in response to action. It's not a new idea humans invented recently. It's the oldest idea in biology. And that holds even as the headlines keep coming from the other direction. OpenAI's GPT-6 Astra and Anthropic's Claude Fable 5.1 both landed this month, and both are really impressive leaps in language, reasoning, and code. Neither one touches the problem this article is about. Getting better at predicting the next word is not the same skill as understanding what happens when a foot meets a wet surface, and no amount of benchmark progress on one closes the gap on the other.

Every organism with a nervous system runs some version of Sense - Reason - Act. A gecko senses the surface in front of it, its nervous system predicts what will happen if it commits its weight to that foothold, and it acts. A fish senses a pressure change in the water, predicts where a predator's mouth will be a fraction of a second from now, and turns before the strike lands. None of that is reflex alone. It's prediction, built on an internal model of how the physical world behaves, refined over an evolutionary timescale so long it's hard to picture. The AI research world model concept, encoding the environment, learning how it changes, rolling the model forward to imagine what happens next, is evolutionary biology's oldest trick, being reverse-engineered in silicon.^[7]

It's not just what these models do that mirrors biology. It's how they get trained. A foal is up and falling and up again within its first hour of life, building its own working model of balance through repeated failure, long before it can run. A kitten pouncing on a piece of string it has no intention of eating is doing the same thing: rehearsing an internal model of how a moving target behaves, so the real hunt goes right the first time it matters. That's the training loop behind something like BeyondMimic: fail repeatedly, refine the internal model, try again, until the prediction holds up. Ha and Schmidhuber's original 2018 paper even called the process "dreaming." The agent trains almost entirely inside its own imagined version of the world rather than the real one.^[7] That's a strange thing to name a training method after, until you remember what a brain is doing every night during REM sleep: replaying the day, without needing to physically re-live any of it, to make the model a little better by morning.

That's not a metaphor I'm reaching for to make a nice paragraph. It's the actual argument for why biomimicry belongs at the centre of the Physical AI conversation rather than at the edges of it as a curiosity about gecko feet and cartwheeling robots. Nature already ran the experiment on what a working Sense - Reason - Act loop needs to look like to survive in an unstructured, adversarial, physical environment. Physical AI is trying to build the same thing on a research timeline instead of a geological one.

7. The Honest Limits

None of this means nature's answer is automatically the right one for an engineering problem. Nature optimises for survival and reproduction, not for whatever a human engineer needs. A bird's wing is a masterclass in aerodynamics and a terrible template for a commercial jet, because a bird never needed to carry 300 people at altitude. Copying form without understanding the underlying function is the most common trap in the field, and Beijing was a live demonstration of exactly that gap: a robot that can now out-sprint any human on earth still can't reliably stop, fall, or pick up a bean the way a toddler can, because a straight sprint down a marked lane is a narrow function to copy, and graceful, adaptable, damage-tolerant movement is the thing evolution spent 3.8 billion years optimising for.

The really useful direction, and the one AI is now accelerating, is using generative design and evolutionary algorithms to search a design space the way evolution does: through iteration, variation, and selection against a fitness function, except compressed from geological time into a compute cluster running overnight. That's biomimicry and AI working together in the other direction: not AI copying a biological mechanism, but AI running evolution's own method faster than evolution ever could.

Closing Thought

There's a small, active community doing exactly this kind of work here in New Zealand. Biomimicry New Zealand has spent the last several years connecting architects, engineers, and material scientists around bio-inspired design, and Te Pūnaha Matatini, the country's complex systems research centre, describes part of its own mission as modelling an entire country as a complex system.^[8] That's a nice echo of an idea I've spent a good part of this year writing about from the AI side. Nature was doing complex systems modelling, at country scale and well beyond it, a very long time before any of us had a word for it. Physical AI's job right now isn't to out-invent 3.8 billion years of R&D. It's to actually pay attention to it.