How AI is quietly accelerating battery and actuator breakthroughs
Part 2 of a five-part series on how humanoid robots quietly became investable while everyone was watching large language models.
A Unitree H1, running a 0.85 kWh battery pack, currently manages about four hours of static operation before it needs recharging. And if it's not vaporware, a Tesla Optimus Gen 2, on a considerably larger 2.3 kWh pack, taps out after roughly two hours of actual walking work. Those numbers matter because they expose the real bottleneck in humanoid robotics. And it's not intelligence in the conversational sense everyone associates with AI, but energy density and joint hardware, two very unglamorous engineering problems.
In the first article of this series, I argued that the commoditization of humanoid robots is the product of several technology curves converging quietly over a decade, and this piece takes apart the specific curve that surprised me most: AI-accelerated battery technology.
Philippe MEDA
If you missed it, Part 1 in this series.
AI is pivotal for unlocking humanoid robots, not the way you think
Solid-state battery energy density has increased from roughly 250 Wh/kg for conventional liquid lithium cells to double that for the new solid-electrolyte chemistries that key suppliers (CATL, CALB, and Samsung SDI) are now sampling for robotics customers.
Battery chemistry research has historically been slow and combinatorially brutal work, testing electrolyte formulations, interface coatings, and thermal stability characteristics across a search space too large for any human research team to cover exhaustively. When I started reading more about it, it reminded me of the gruesome, kitchen-sink approach to protein folding we had in the late nineties. Which was, testing stuff, a bunch of stuff, as much stuff as possible, and praying for a random (I mean... stochastic) positive output. This is exactly the kind of problem where machine-learning-assisted materials discovery earns its keep. And yes, screening candidate compounds computationally before anyone touches a lab bench has begun to compress years of iterative laboratory work into months.
Gordon Moore's original observation, that the number of transistors on a chip doubles roughly every two years, has no real equivalent in battery chemistry. And Lithium-ion energy density has historically improved at something closer to 3-4% a year since Sony commercialized the first cells in 1991. But what used to be an eighteen-to-twenty-year lithium-ion doubling cycle has become roughly five years, with gains closer to 15% annually. If we add current AI exploration efficiency to this, we can put the actual efficiency zone of a robot running eight hours without a battery hot-swap somewhere between 2029 and 2031.
What generative AI contributed to this story isn't a smarter robot brain but a faster search across an overly complex and unreliable chemistry space. Ironically, such an acceleration is arguably a more consequential outcome of the current AI wave than anything a large language model has produced for the consumer market.
Actuator manufacturing, the other silent bottleneck AI is also helping to break
Another major bottleneck for humanoid robots is simply mechanics. Today, servomotors and joint assemblies still account for roughly 50% of a humanoid robot's total unit cost. Which is why the actuator supply chain (not the software stack) has been the more stubborn constraint on scaling humanoid production.
Fourier's harmonic reducers only recently entered mass-production testing, and Unitree's self-developed M107 joint motor is only now clearing the mass-production threshold that has held back the entire category for years. Both developments depend on simulation and optimization techniques that lean heavily on the same machine-learning infrastructure originally built for language and image models, repurposed to model mechanical stress, wear patterns, and torque efficiency across thousands of candidate designs faster than physical prototyping ever allowed. Obviously, none of this is as visible or as easy to demo on stage as a chatbot answering a question. Which is precisely why it hasn't received any press attention. That, and let's be honest... What a boring topic, right?
And yet, there is a broader implication here for how large European corporations should be reading the current AI cycle, one I keep coming back to with clients across the sectors I work in, from automotive to aerospace.
The public conversation about AI has been dominated almost entirely by language models, and that framing quietly encourages strategy teams to treat AI as a productivity or content tool rather than as a research accelerant embedded inside adjacent fields like chemistry, materials science, and mechanical design. But this is the real connective tissue of the innovation field! A defense group or an automotive tier-1 supplier evaluating its AI exposure purely through the lens of copilots and chatbots is missing half the story, where AI-assisted search is compressing R&D cycles in exactly the unglamorous domains – battery chemistry, actuator design, alloy formulation – where (ironically) their own competitive advantage has traditionally lived.
That is not a hypothetical risk.
It is already visible in how fast the humanoid robotics supply chain has moved from prototype to near-commercial scale in under two years, a pace that would have been implausible under the old materials-science research cadence.
The next article in this series takes the argument in a more provocative direction, asking why the human body, a genuinely mediocre piece of industrial engineering by almost any measure, turned out to be the right chassis for AI to learn from in the first place.
Again, stay tuned!
If you missed it:
Philippe MEDA
