I've spent years in China building robots. I can tell you this playbook won't work: Pre-seed $: Invent the ultimate actuator. Seed $$: Figure out mass production. The hard part isn't inventing a beautiful topology. It's process engineering. The real delta between the US and China is the ability to execute the enormous amount of physical x mental work required to iterate thousands of times on product designs and manufacturing processes. Partly because much of the tacit knowledge / management systems has left the US, but also because of the sheer human labor required. This isn't something you can solve by buying traditional automation lines either. It requires: 1. Ideation driven by human intuition. Walking up and down the line, talking to assembly workers, and developing hypotheses about what and where improvements can be made. Raw human hours and attention spent understanding a specific manufacturing process. 2. Large batches of experiments across multiple variables. Coordinating experiments that can easily involve 50+ people, issuing instructions, and tracking the execution and results of each experiment. 3. Purpose-built test design to identify key metrics. For example, a change in gear hobbing profile might require accelerated aging tests on a dyno, while a change in bearing tolerance might require NVH testing. 4. Data + human intuition-driven analysis. Interpreting test results and deciding whether to implement a process change or run the next set of experiments. 5. Strict process control. Reissuing work instructions to the line and ensuring every process is followed to the letter. The best manufacturers in Asia, from Foxconn to TSMC, have built their competitive advantage on systematically turning thousands of physical experiments into better manufacturing processes. Their moat isn't just production scale. It's the organizational machinery that makes manufacturing learn and improve faster than anyone else. I don't think most people in SF clock this. Say you're manufacturing actuator gears in-house in the US. After you've made your first article, will you hire a team of 10 process engineers to coordinate the 50 technicians running concurrent experiments with designs, processes, and materials? To understand wear patterns across every pin in a cycloidal drive after 10,000 hours? To reduce backlash from 15 arcmin to 5, then to 1? Just to keep pace with China's progress over the next few years? Do you have the skill set and management systems in place to do this? That's why our approach isn't just to automate manufacturing. It's to automate the process of learning how to manufacture better. Automate simulation and design optimization. Automate assembly and physical testing. Automate the design and execution of process experiments. Feed real-world experimental data back into simulation to continuously improve sim-to-real accuracy. The goal is to turn what currently takes 50 engineers coordinating thousands of experiments into a closed-loop system that continuously learns and improves. Then the same robotic line mass-manufactures the product. PS: Coming back to actuator design, many actuator designs are already commoditized in China. The project in the video is open source, and CNC machining cost in China is just $27 (watch the video). I hope people understand this reality. I'll link more examples in the replies.

If I was 21yo with free time right now: I'd vibe code AlphaKickbox: humanoid kickboxing environment in Mujoco Full contact physics, torque control, multi-agent env And let lightly-pretrained rl agents self-play until they hit alphazero levels Then publish a second paper on benchmarking existing vlas on it Then publish a third paper on addressing the sim2real gap And make money hosting robot kickboxing tournaments Very venture scalable, flashy, market-making, within the current hype trends

I can’t believe it but I recognize the underlying humanoid from its body language too 🤯 Feels like I accidentally grokked a new language and this is the first time I'm understanding someone speaking it in the wild

Hosting a houseparty (for nerds) this Sunday for friends from top robotics / frontier AI labs and friends applying it to reindustrialize America. Come hang out! luma.com/5794f6gd

These guys are really forward forward deployed @makermodsai @IsaacSin12 @Ryan_Resolution 🤯 Thanks for the unboxing experience time to benchmark tf out of you 😊

Humanoid hardware is like iPhone 1 right now. Shitty but usable by enthusiasts 2026-2028 we will get to IPhone 4 level. Usable but lowkey missing ecosystem and software (intelligence) polish By then low-hanging fruits like overheating on regular use, unreasonable inaccuracies/jitter will be solved. It no longer bottlenecks basic deployment. From then onwards it’s a loooooong tail of intelligence, hardware, ecosystem improvements. Its going to be a marathon.

We used to run RL on Countdown (4 number version of this) and debate why it works - is it amortizing search - just exploring every probability and memorizing it first - if it is, if we do that in a very general way isnt that just what intelligence is? Crazy to see it in person, crazier to see frontier models ace it so quickly

Thinking automated factories is automated assembly is like thinking a restaurant is a place where waiters bring you food. What systems will you build to improve quality and continue innovating new dishes. What internal tools and tacit knowledgebase will you build, to enable talent to leverage it to the maximum for faster and cheaper innovation of future dishes. How you build the supply chain and partners, and push everyone to innovate with you to make better dishes. I'd be super happy to see some new-age US manufacturers become experts at one dish, and then diversify to many dishes within the same distribution To say that you are going to be doing cooking-as-a-service for many dishes from different cuisines from day one Because - it's better to not own the design loop - so don't actually know why, what recipe - just be experts at executing recipe - because then you're free from market risk, you just have to cook whatever recipe people pay you for - for everyone, across all recipes from various cuisines - when you're starting from zero tacit knowledge - but that's ok becuase tacit knowledge from Chinese cuisine will generalize to Italian cuisines, because the surface-level actions of frying, flipping are generalizable - these actions aren't even tacit knowledge that's physical intelligence that will be commoditized in the next few years anyways how are sf vcs buying into these stories is a mystery to me

Any robotics company in SF that has PiperH than we can rent We're benchmarking arms!

China's ecosystem is running at different speeds, different ecosystem, next levels of low margins. - Best supplier (#1 out of 6) has 13 day lead times - Bottleneck is grinding tool lead time of 6 days, wants to get total lead time down to 9 days. - 190+ iterations for 'the biggest vacuum cleaner company' (probably Dreame) in the last 6 months, and cuts 3115 drone motor magnets as commodity - For reference our NDAA-compliant vendors lead time is 6 weeks, US/EU players are 6 months. All this pressure exists because they're ecosystem is competitive. They hyperoptimize the management of their factory from end to end, pressure every ring of their suppliers, employees, process and design engineers to stay competitive against other magnet makers. And you will discover the why and how over tea and cigarettes while they watch your face to guage your scale, seriousness and trustworthiness

If I thought I could change the course of manufacturing by making actuators, I would have - Raised 1mil, built a team of 15 with my best bros in Shenzhen to iterate twice a week - Raise Seed to buy automated line for Q3 2027 Just in time for China make a better, cheaper actuator

SuperTrack (2021) -> SuperDyno (2025) -> SuperMimic (2026) Why'd it take 5 years to go from sim to real? Incremental improvements through inspired experiments. Premise: World model learns from fresh simulator data Actor (policy) learns from world model 1. Train one world model on robot states (not pixels), shared across motions 2. At every control step: observation + goals (reference motion) → [policy] → actions 3. During each actor update: - Freeze the world model’s weights - Run 'BPTT' through 'imagined rollouts' to train the actor Imagined rollouts: [ actor → frozen world model → actor → frozen world model → ... ] Backpropagation through time (BPTT), in SuperMimic: - unroll imagined rollout - backpropagate discounted imagined rewards + terminal critic value through the loop - compute gradients with respect to policy weights - update policy weights What did it take to get from simulated characters to a physical humanoid at this scale? Every paper brings its own very well curated bag of tricks. Pre 2022 it sounded ridiculous to me to train any of this. One model needs to learn simulator dynamics Then one model learns to control a robot within that model?? My SAC can barely work in a task that takes 32 timesteps bro The obvious approach runs into exploding/vanishing gradients, difficulty learning contact dynamics, and discontinuities from termination. It took a whole series of papers, each bringing a carefully curated bag of tricks to inch us here. Some landmark papers: SuperTrack (2021): 6.5 hours of motion, validated in simulation. 1. Use root-relative states to remove irrelevant position and heading. 2. Train on multi-step windows: 8 steps for the model, 32 for the policy. Enough to lookahead but not enough to destabilize training. SuperDyno (2025): ~40 hours of motion, validated in simulation. 1. Use cycled synchronous early termination during simulator data collection. SuperMimic (2026): 264 hours of motion, selected motions demonstrated on a physical Unitree G1. 1. Add a terminal critic to account for return beyond the rollout. 2. Use soft termination probabilities to keep imagined returns differentiable. 3. Randomize physics and condition the model on those parameters; give the actor noisy sensor observations for hardware transfer. 4. Favor difficult motion segments when resetting episodes. Each of them didn't come up with these tricks on the first go They're results of experimental cycles spannning across research assistants, Phds, post-docs, collaborators, and now agents too. You are seeing the results of what made the final cut Cool stuff! Funky to see learned dynamics models getting the 'world model' branding though boooooo

We're not going to keep pace with China's actuator development by doing process engineering with humans. China isn't going to reach actuators with precision that can perform surgery on cells with their current playbook of doing process engineering with humans. I believe everyone will move towards automated design and process improvement - the question is just how. Let me give you a real world example of how this works first I can design our Caldera-0-3115 motors to have an airgap of 0.08mm airgap instead of 0.16mm airgap but I can't manufacture it. If i could, this would translate to about 8% max torque at 2.5kW max throttle - which Counter UAS folks would absolutely love. This is more impactful than any design improvement I can make from optimizing EM design with Ansys and maaaybe set a new world record. But I just can't manufacture it - even if I did, it would die within x hours of flight time because at some point the stator and magnets would start colliding. Because my current manufacturing process creates error tolerance stackup from 1. flipping/refixturing of the multi-step lathing process of the rotor bell not perfectly coaxial 2. press fits 3. adhesives 4. tolerancing the grinding of magnets 5. at 22k rpm, the remaining 0.x miligrams of imbalances after dynamic balancing that sway the bell from one side to another as well which requires some margin. 6. warping from repeated thermal expansion I can't improve them without CNCing with different tools / cutting operations / fixtures; try different parameters for press fit, try different adhesives which responds differently to thermal expansion, etc. And that's just the airgap. Every feature has its own quirks and optimizations to be made. And this isn't a one time thing, this is a marathon that all actuator-makers are running, always, for every SKU. With the added complexity of a gearbox. So I understand Why @atlasmotion does custom motor design + US prototyping then does volume production in the Philippines. Their experienced manufacturing labor gives you the flexibility to handle different products AND process experimentation. Why @parthingle_x is building custom CNCs specialized for motors/actuators and a factory knowledge base with, telemetry and scheduling capabilities. This makes sense to me because it's a specialized tool to automatedly improve to cut better parts for motors. Why Almetra capture live factory floor video, immediately convert raw human movement into structured digital work steps, and orchestrate real-time floor feedback that systematically drives automated process improvements. We at Caldera are going the route of fully automating design and process iteration via factory intelligence, robotic assembly and software integration over all machines that participate in actuator manufacturing. All of us are building pieces of the same capability: running more experiments, learning from them faster, and getting the improvements into production. That's what US manufacturing needs to compete. The motor you can build today is the starting point. What matters is how quickly you can build a better one. Please tag people with a blurb of what they're doing if you think they are working in this field as well! It'd be super interesting to see what everyone is building and what are their takes.

Told @JasonMa2020 @DynaRobotics VLAs bad for manufacturing because precision doesn't scale with high quality data, it just teaches it to be more generalizable. and he said 'im not saying its a skill issue but...' kill me please He's right - he's saying turn precision problem into a retry and recovery problem that VLAs are good at. Humans cannot one-shot the same task either. We feel around with touch and vision until it works. Like plugging in a USB or inserting a shaft. Very light pressure around where things should go, to give enough signal to tactile fingers @wuji_global / visual deformation of grippers @GeneralistAI / joint torque @UnitreeRobotics So obvious in hindsight. Time to re-collect trajectories for training @HuanbinChe0dde

We're on the same mission