Image credit: Lottie animation by Centre Robotics (LottieFiles Free, used with credit). · source
Most of the coverage I've seen about AI-designed robots misses the point entirely. The headlines focus on the flashy stuff, robots that look weird or move in unexpected ways, but they skip over the actual engineering problem that's kept this field stuck in neutral for years. Two new papers from arXiv tackle that problem head-on, and if you're not paying attention to this work, you should be.
The problem is simple to state and brutal to solve: how do you train a robot's brain when you're also changing its body? Every time you tweak the morphology, the old controller becomes useless. It's like teaching someone to ride a bike, then handing them a unicycle and expecting them to just figure it out. Multiply that by thousands of design iterations and you've got a computational nightmare that makes most co-design approaches impractical outside of simulation.
I've seen this movie before. Back in the early days of neural architecture search, we had the same scaling wall. Train a new network from scratch for every architecture? Prohibitively expensive. Use one giant network for everything? You get mediocre results across the board. The solutions that eventually worked involved modular approaches and knowledge transfer, which is exactly what these two papers are exploring for robot bodies.
The first paper, ECo-MoE, comes from researchers who've clearly thought hard about the modularity question. Their approach uses a mixture of experts, basically a collection of specialized neural modules that get activated or deactivated depending on what kind of robot body you're dealing with. Different body plans trigger different combinations of these learned sensorimotor circuits. The clever bit is that when evolution produces a new species of design, you can overhaul one part of the controller without destroying the knowledge stored in the other modules.
Related coverage
More in Research
Researchers are building systems where robots evolve their own bodies and controllers together, then save what they learn for next time.
James Chen · 1 hour ago · 6 min
Three papers crossed my desk this week that suggest we're finally getting serious about making robots do what we actually tell them to do.
Robert "Bob" Macintosh · 3 hours ago · 4 min
Researchers are finding ways to train robots with far less data, using human corrections and physics simulators instead of millions of demonstrations.
James Chen · 3 hours ago · 6 min
A batch of new research papers suggests we might finally be solving the sample efficiency problem that's plagued robotics for years, and I've seen this inflection point before.
This isn't revolutionary in the sense that mixture of experts is a well-established technique in machine learning. But applying it to embodied evolution in a way that actually preserves ancestral knowledge? That's harder than it sounds, and the results suggest it works. They also introduce something they call "evo by demo," where you can plug in pretrained expert policies to steer evolution toward specific morphological traits you want. Think of it as giving evolution a nudge in a particular direction rather than letting it wander randomly through design space.
The second paper takes a different angle, and honestly it might be the more interesting one.Auto-Robotist addresses what I'd call the institutional memory problem in robot design. Most evolutionary approaches are memoryless in a frustrating way, the simulator results influence the next generation but nobody writes down what was learned. It's like a company where employees leave and take all their knowledge with them, over and over, forever.
Auto-Robotist uses large language models to distill the search process into an explicit skill library written in natural language. Each skill contains a structural archetype, rules about what works and what doesn't (with evidence!), and the actual evaluated designs that support those conclusions. When the system searches for new designs, it retrieves relevant skills to guide the process. After evaluation, it updates the library through add, diagnose, and merge operations.
The researchers tested this across seven tasks in EvoGym, covering locomotion, traversal, and object interaction. The results show it improves cold-start search and, more importantly, transfers learned skills to larger design spaces. That transfer capability is the key thing here. You're not starting from zero every time.
Now, call me old-fashioned, but I'm skeptical of anything that relies heavily on LLMs for core functionality. The failure modes are well-documented and the brittleness is real. But what I like about this approach is that the skill library is inspectable. You can actually read what the system thinks it learned and check whether it makes sense. That's a meaningful improvement over black-box approaches where you just have to trust the neural network knows what it's doing.
The bigger picture here is about making robot evolution practical at scale. Right now, designing robots through evolutionary methods is mostly an academic exercise because the computational costs are prohibitive and the knowledge gained doesn't transfer well. If you can solve those problems, you open up the possibility of automated design pipelines that actually work in industry settings.
We're not there yet, let me be clear about that. Both papers are working in simulation (EvoGym specifically), and the gap between simulated results and real-world performance remains substantial. The ECo-MoE paper doesn't provide details on how many training iterations are required, which matters a lot for practical deployment. And Auto-Robotist's reliance on LLMs introduces dependencies on models that are themselves rapidly changing and not always reliable.
But the direction is right. The field has been stuck on this co-design scaling problem for a long time, and these papers represent genuine progress on the fundamentals. Not the flashy, look-at-this-weird-robot kind of progress that gets clicks, but the boring infrastructure work that eventually makes the flashy stuff possible.
I've been covering tech long enough to know that the important advances usually look mundane at first. The kids building these systems are solving the right problems, even if the solutions aren't perfect yet. Modular controllers that preserve knowledge across morphological changes. Design memories that make search results reusable. These are the building blocks.
What remains unclear is how quickly this translates to real hardware. Simulation-to-reality transfer is its own can of worms, and neither paper addresses it directly. The computational requirements, while improved, are still substantial. And there's the question of whether these approaches generalize beyond the relatively simple task domains they've been tested on.
Still, if you're trying to understand where robot design is actually heading (as opposed to where the press releases say it's heading), these two papers are worth your time. The hype cycle around AI-designed robots will continue regardless, but the real work happens in papers like these, one unglamorous problem at a time.
If you want to argue about any of this, my email's on the about page.