Image credit: Lottie animation by Centre Robotics (LottieFiles Free, used with credit). · source
Two papers dropped on arXiv this week that, taken together, sketch out something I've been waiting to see for years: robots that can evolve their own morphologies while actually retaining useful knowledge between generations.
Let me be clear about what this is and isn't. These are simulation results. No physical robots were harmed, or even built, in this research. But the underlying ideas address a real bottleneck in how we design robotic systems, and the approaches are different enough to be worth examining side by side.
If you want to optimize both a robot's body and its controller simultaneously (what researchers call "co-design" or "co-optimization"), you run into a scaling wall fast. The naive approach is to train a separate neural network controller for every single body design you evaluate. From my time building hardware, I can tell you this is roughly as efficient as it sounds, which is to say, not very.
The alternative has been to train one universal controller that works across all possible body configurations. This sounds elegant until you realize it produces conservative, mediocre designs. The controller has to hedge its bets across every possible morphology, so you end up with robots that are acceptable at everything and excellent at nothing.
Both papers this week try to find a middle path. They're asking: can we preserve what evolution learns without starting from scratch every time?
The first paper, ECo-MoE, comes from a team that's essentially treating robot control like a committee problem. Instead of one monolithic controller, they use a mixture of expert neural modules. Different body configurations activate different combinations of these experts.
Related coverage
More in Research
Forget the hype about AI designing robots. These researchers are solving the boring, fundamental problem that's held back the field for decades.
Mark Kowalski · 1 hour ago · 5 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.
The technical setup: they co-optimize a distribution of latent design vectors (the genotypes) alongside this mixture of control experts. Each decoded robot design (the phenotype) gates which experts fire based on its position in latent space. So a legged design might activate one set of sensorimotor circuits, while a wheeled design activates different ones.
What I find interesting here is the preservation of ancestral knowledge. When evolution produces a new species of design, you can overhaul one part of the controller without destroying what other expert modules have learned. It's modular in a way that actually matters for iterative development.
The researchers also introduce what they call "evo by demo," where you can plug pretrained expert policies into the mixture to steer evolution toward specific morphological traits. Want your evolved robots to have legs? Give it an expert that already knows how legs work. This seems like it could shortcut a lot of the random flailing that characterizes early evolutionary search.
The paper claims this provides a "scalable alternative" to existing approaches, though I'd note they don't provide direct computational comparisons with baseline methods. That's an ambitious framing without the numbers to back it up. The real test would be seeing this run on more complex morphology spaces than the demos suggest.
The second paper, Auto-Robotist, takes a completely different approach. It's using large language models as proposal generators for evolutionary design, which isn't new. What is new is that they're actually saving what the system learns.
Most LLM-driven design loops are, as the authors put it, "memoryless." The simulator evaluates designs, results shape the next population, and then all that hard-won knowledge vanishes. Auto-Robotist instead distills morphology search traces into an explicit natural-language skill library.
Each skill stores:
A structural archetype
Evidence-grounded positive and negative rules
The actual evaluated designs that support those rules
So instead of implicit knowledge buried in a population distribution, you get inspectable, auditable design principles. The system can retrieve relevant skills to condition future LLM edits while still maintaining a genetic algorithm mutation path for exploration.
The results across seven EvoGym tasks (locomotion, traversal, object interaction) show improved cold-start search and, more importantly, successful transfer of learned skills to larger design spaces. When they moved from 5x5 to 10x10 design grids, the skill-conditioned transfer outperformed baseline genetic algorithms on every task.
Look, I'm inherently skeptical of LLM-based robotics claims because the field has produced a lot of impressive demos that don't survive contact with physical hardware. But the memory architecture here addresses a legitimate gap. The fact that you can inspect why the system made a design choice is valuable for debugging and for building trust in automated design pipelines.
Here's where I have to be honest about limitations. Neither paper provides the kind of computational cost comparisons I'd want to see. How many GPU-hours does ECo-MoE require versus training individual controllers? What's the inference overhead of the mixture-of-experts gating? The Auto-Robotist paper mentions they'll release code upon acceptance, but acceptance hasn't happened yet.
The Auto-Robotist team does show transfer results, which is something:
Task Type
GA Baseline
Skill Transfer
Locomotion
Baseline
Outperformed
Traversal
Baseline
Outperformed
Object Interaction
Baseline
Outperformed
They claim improvement across all seven tasks, though exact performance deltas aren't specified in the abstract. That's the kind of thing you'd want to see in the full paper.
From my time in hardware engineering, I've watched companies spend months hand-tuning robot designs for specific tasks. The promise of automated co-design has been around for decades, but the practical barrier has always been sample efficiency. You can't afford to evaluate thousands of physical prototypes.
These papers are both, in different ways, trying to make the search process more efficient by preserving and reusing knowledge. ECo-MoE does it through modular neural architectures. Auto-Robotist does it through explicit linguistic memory.
Neither approach is ready for production deployment. We're still talking about simulated environments with simplified physics. The gap between EvoGym and a factory floor remains substantial. But the direction is right.
The question I keep coming back to is: what happens when you combine these approaches? A mixture-of-experts controller that's guided by an LLM with explicit design memory? That's probably the next paper, and I'd actually read that one closely.
I should note what neither paper addresses, which is sim-to-real transfer. Evolving robots in simulation is one thing. Getting those designs to work on physical hardware is another thing entirely. The actuator dynamics, material properties, and manufacturing constraints of real robots introduce failure modes that no simulator fully captures.
It's too early to say whether these memory-based approaches will help or hurt sim-to-real transfer. You could imagine that explicit design rules ("don't make legs thinner than X") would encode physically meaningful constraints. Or you could imagine that the rules learned in simulation would be precisely wrong for reality.
We don't know yet. Both teams would need to demonstrate physical builds before I'd get genuinely excited.
Both papers suggest their approaches are generalizable beyond the specific tasks they tested. ECo-MoE hints at applications to "freeform evolution toward canonical structures." Auto-Robotist claims their skill library could convert "expensive physical evaluations into reusable, auditable design principles."
These are, sort of, promissory notes. The proof will be in follow-up work that actually tests generalization and, eventually, physical deployment.
For now, what we have is two research groups independently recognizing that evolutionary robot design needs memory. That convergence is interesting even if the specific implementations are still early-stage. The field has been stuck on memoryless optimization for a while. Maybe this is how it gets unstuck.