Midcentury

Data and Simulation Infrastructure for Physical AI

[Midcentury @MidcenturyAI]

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

[Midcentury @MidcenturyAI]

RT @ckxyz_: 10m+ hour robotics datasets are about to be normal

the kicker is frontier labs will only care about usable hours. sub-scale p…

[Ali @robot_xyz_rpy]

“As more data becomes available, the bottleneck is shifting to data quality.”

We’ve 2M+ hours of ego data at @MidcenturyAI . With Sortex, we’re making that corpus searchable and curating it around specific robot training needs.

Enjoy reading Sortex below

[addeagupta @addeagupta]

We built Sortex because “2M+ hours” tells you almost nothing by itself. What matters is what’s in there: whether each clip actually shows the work clearly, which tasks and environments the corpus covers, and what every new recording adds. Our goal is to make quality measurable and inspectable at both levels, from a single frame up to the whole dataset. Read more below!

[Midcentury @MidcenturyAI]

Sortex was built by Addea Gupta (@addeagupta) and David Guo (@dgmonsoon). Read more below.

https://t.co/TVQtnrhTAb

[Midcentury @MidcenturyAI]

Sortex: our data quality platform built to search and filter the world's largest egocentric dataset.

Across 50+ environments and 20k+ tasks, it uses an in-house model and evaluation harness to score each clip's usability, selecting for specific training objectives by:

→ Narrating tasks and steps
→ Checking hand and action visibility
→ Comparing against the rest of the corpus

We also tested Sortex on popular open datasets and found they differ widely in quality and coverage.

[Midcentury @MidcenturyAI]

We also ran Sortex on random clips from three open egocentric datasets.

About one in six clips failed screening in EPIC-KITCHENS-100 and RekaDaily-10k, mostly for darkness, shake, blur, and faces.

In Gen-HumanEgo, 42% of the clips that passed were near-duplicates of another clip in the sample.

[Midcentury @MidcenturyAI]

Frontier hand pose now costs 41% less than it did a few weeks ago.

Our team has been shipping updates to MC-EgoHands nonstop: model loading, batching, video encoding, validation and export recovery. We’ve also sped up hand-pose delivery by 34%.

Each reduction in compute per clip moves us toward the millions of high-quality hours robots will require.

[Ali @robot_xyz_rpy]

Quick MC-EgoHands update:
---------------------------------
Since release, I’ve been working on improving processing costs and hand-pose delivery speed. Here’s a few of the optimizations:

> Keep GPUs busy: Inputs are prepared while inference runs, and we batch up to 64 hand crops per call, so the GPU spends less time waiting.

> Run more in parallel: Workers and cached weights are reused across clips, and validation and packaging run off the inference GPU so it starts the next clip sooner.

> Stop repeating work: We reuse RGB verification results, keep the original video when it already aligns with the poses, and recover failed exports from saved predictions.

> Same outputs: On the tested clips, smoothed hand outputs are bit-for-bit identical, with frame alignment and LeRobot v3 exports verified.

[DG @dgmonsoon]

neural rendering is at the core of our strategy. Early evals show it's likely the best solution available

DM me if you'd like to try!

[Midcentury @MidcenturyAI]

@solidSF @B_S_N_Y but wait there's more

https://t.co/aqKIG75BXc

[Midcentury @MidcenturyAI]

Introducing Akari: a frontier learned rendering model for robotics simulation.

Akari generates photorealistic rendering for any physics engine to close the sim-to-real gap.

Trained in-house, it is the fastest and highest-fidelity learned renderer in any sim platform: 0.046 LPIPS at 0.5s/frame, consistent across 1K+ frames.

[Midcentury @MidcenturyAI]

@cspaliwa1 you can join the early access waitlist here: https://t.co/hZHUKFtrOd

[Midcentury @MidcenturyAI]

@robot_xyz_rpy @i2rt_robotics 👀

[Chetan Kulhari @ckxyz_]

sim2real gap is one of the biggest bottlenecks in robotics. we trained akari in-house, the first of the major model breakthroughs that will close it

turns out a data moat and a cracked research team can compound pretty fast

[Brandon Samaroo @B_S_N_Y]

There are very few teams out there with the resources and research talent to tackle learned sim. We are adamant that solving the sim2real gap will require neural systems to model the increasingly complex laws of the universe— and Akari is the first step towards that bet.

Learned sim is the way. Introducing Akari.

[Midcentury @MidcenturyAI]

Akari is a proprietary Midcentury model and will be available exclusively through our simulation platform, Matrix.

Development of Akari was led by Brandon Samaroo (@B_S_N_Y). Full report below. https://t.co/OGKvVVIBhl

[Midcentury @MidcenturyAI]

Instead of calculating how light bounces through a scene to produce shadows, reflections and material detail, Akari learns what those should look like from data.

As a result, Akari isn’t bounded by compute or hand-written code, and its rendering quality scales directly with training data.