Introducing Midcentury.
We’re building the data and simulation infra for physical AI.
Today, we’re coming out of stealth with a $15M Series Seed to scale robotics beyond polished demos.
We’re already supporting frontier labs with:
→ The world’s largest egocentric dataset: 2M+ hours, 50+ environments, 20,000+ tasks
→ Matrix: a frontier simulation platform scaled with our real-world data to evaluate and post-train policies at scale
Thanks to everyone who joined our CoRL happy hour, and to our speakers from @SkildAI, @physical_int, @amazon, and @nvidia.
Simulation was everywhere this year. A few takeaways:
1. Astra collapsed the cost and time of real2sim. Environments, assets, and URDFs that took months now take days. The bottleneck has mainly shifted to sim2real, getting the physics of contact-rich manipulation like friction, mass, and deformation accurate enough to train on
2. The leading approaches to closing the sim2real gap: domain randomization to cover what sims get wrong, SysID to tune the physics closer to reality, and neural simulators that learn dynamics from real data
3. Tactile is still hard. Hardware isn't standardized, data is inconsistent and expensive, and there's no consensus on what works. There's growing interest in scaling tactile in sim because the real world alone is unlikely to yield the diversity and scale dexterity needs
4. Long-horizon is the new deployment frontier. The standout demos ran full multi-step tasks zero-shot and that's where memory and error recovery start to matter more than raw success rate
5. Evals are lagging. Real-world testing is slow, public benchmarks aren't trusted, and success rate misses what actually matters for deployment, like speed and recovery from failure
6. Teams are writing their own physics solvers. Skild built Steel and more labs are going the same way instead of relying on existing engines
sims working is the best case for robotics data labs, not the worst
sim will unlock the superintelligence moment for physical ai but it's not sim vs real. real world data like ego and tactile grounds sim in the right physics and distribution and sim scales it from there
AVs already did this. waymo built simulationcity on top of 20m+ real miles, recreating the rare encounters its fleet actually saw and running billions of variations in sim
LLMs are the latest proof. RL environments are carrying frontier post-training and they've made labs way hungrier for real workplace data. reliable sim starts with real world data
“Almost nobody has stereo off the shelf.”
At CoRL this week - we’ve built one of the largest OTS stereo datasets and benchmark it obsessively.
Would love to talk stereo, data quality, and post-training results. DMs are open.
CoRL is unreal this year. our rooftop happy hour got ~10x more signups than the space can hold
would love to meet more researchers and builders at the frontier. dm me to chat tactile data, simulation, deployment, or research partnerships
real2sim (and most of sim for that matter) is one of those low floor/high ceiling skills— anyone with an Astra subscription can participate, but it takes real technical depth to be at the top.
when you are training perception-heavy models like RFMs, every detail matters.
we’re renting out a rooftop at CoRL next Monday for a happy hour
a stacked room of researchers from NVIDIA, Physical Intelligence, Skild, Dyna + more working across robotics data, simulation, and deployment
few spots left, join us!
luma.com/u6rr5lql
learned models supervised on physics-rich data are the best bet for closing the sim2real gap
data is still the bottleneck. get enriched ego data with tactile and depth, at scale and diversity, and you can solve sim for dexterous, contact-rich manipulation
good to see this direction picking up. it's been core to our thesis, more soon!
the best robotics models will come from teams that understand failure modes best and know exactly what data and supervision it takes to fix them
three failure modes stand out here: grasp selection, 3D positioning, and failure recovery
our post-training experiments are already showing meaningful gains from more targeted data and supervision. if you're curious what we're finding so far, reach out!
This halloween @archiexzzz , @notnoahphan and I were testing 4 models on a pumpkin 🎃 @i2rt_robotics YAM gripper can lift only by the stem tip.
TL;DR:
> @dreamlabs_ai - Dream-1: tried the body once, switched to the stem, and put it in the box.
> @OpenAI - Astra: aimed for the stem but closed too high. We suspect a depth or positioning error.
> @PokeBot_Global / RoboAPI: kept trying around the body and top without securing a lift.
> @Google@GeminiApp : reached for the stem but never got a grip.
> Failure analysis: grasp choice, positioning, and recovery need to be measured separately.
> Data needed: grasp labels for this gripper, measured 3D targets, and failed attempts paired with successful corrections.
1/5
For anyone curious about why we’re confident this work is worth pursuing:
We tested whether human egocentric data could help adapt π0.5 to LIBERO-10 with limited robot demos. The comparison with EgoDex motivates further testing for a number of reasons. 🧵
The research team has been loving this quick research/experimentation format. It really gives us a way to probe our progress, identify improvements, and share simple learnings with the community quickly— expect a lot more of these in the near future.
Check out our latest one where we put our proprietary (not for long 😉) data up against EgoDex!
This is the first in a series of experiments to better understand what makes human data useful for post-training.
We're continuing to study how data selection, viewpoint diversity, and action supervision affect policy generalization.
Full report by @robot_xyz_rpy below.
midcentury.xyz/blog/shakey
To post-train the full π0.5 model, we converted estimated human hand motion into action targets and mixed the data with robot demonstrations.
Each training arm used the same budget of 20 robot demonstrations, two for each LIBERO-10 task.