Ilya

Robotics data: Ego • Exo • Depth • IMU
Founder @FilemarketAI • Alumni @Techstars
Data Trust infrastructure for Physical AI by 2030
Be Water 💧 一曰慈，二曰儉，三曰不敢為天下先

[Ilya @UnderKong]

Data is the entry point. Robot operations are the market.

Next: our flagship center in Bangkok.

let's build it 🤝 @speedrun @andrewchen @JoshLu @FileMarketAI

[Ilya @UnderKong]

We just applied to a16z speedrun SR008 🧵

FileMarket is building the orchestration layer for physical AI: human data, robot evals and deployment, from our own robot centers.

[Ilya @UnderKong]

Where we are:
36,000 hrs delivered to a top US humanoid company
300,000 more hrs signed
2 frontier labs executing
120 capture rigs built in-house
Break-even

[Ilya @UnderKong]

That's what we're building into our physical AI centers at @FileMarketAI.

Figure's spend makes me optimistic. Its in-house move reminds us labs need a real reason to pick us.

Training robots? What data are you still struggling to find? DMs open 👇

[Ilya @UnderKong]

Figure just spent ~$8.1M on data. In ONE week.

And it's pulling 81 minutes of data every second through its app.

Demand for robotics data is exploding. But here's the part most people miss 🧵

[Ilya @UnderKong]

2/ Harder collection. Synced multi-cam, tactile, teleop demos. That means hardware upkeep, operator training and QA. More collectors won't fix it.

3/ Speed to change. New objects, new tasks, new gear. A setup built for yesterday's brief gets stale fast.

[FileMarket AI Data Labs @FileMarketAI]

🚨 BREAKING: @nvidia drops FLORA — an OPEN-SOURCE robotics engine that’s ~125x CHEAPER than frontier LLMs! 🤖💸

Spots 53% of robot fails, unifies captions + rewards, and labels 200k hrs/day for GR00T.

#NVIDIA #Robotics #OpenSource #PhysicalAI

[Ilya @UnderKong]

@AIgorStadnyk Very interesting point!🤓

[Ilya @UnderKong]

At FileMarket Labs (@filemarket_xyz) , we organize targeted physical AI data collection around these exact failure points—giving teams high-quality demonstrations to adapt and stress-test their models. 🦾

Building robotics models or planning data collection? Let’s talk: https://t.co/0mVf6svMSK

#EmbodiedAI #Robotics

[Ilya @UnderKong]

Simulation Benchmarks (Developer-Reported) 🖥️

• LIBERO: 98.7% success for @Xiaomi Robotics-0, 87.3% for @huggingface SmolVLA (450M)
• RoboTwin 2.0: 92.8% success for Robbyant's LingBot-VLA 2.0
⚠️ Simulation scores reflect isolated internal setups, not a unified leaderboard.

[Ilya @UnderKong]

The Real Bottleneck: Edge-Case Data 🔄

Deploying physical AI into production exposes failure points no simulation catches:
• Recovering from dropped objects
• Handling novel tool geometry
• Dealing with dynamic lighting & slip

Models only go as far as their edge-case data.

[Ilya @UnderKong]

Real-World Benchmark: RoboDojo-RC Tier 1 📊

Independent testing by @robocurve (120 attempts, mean task progress):

• Dream-1: 59.3% • MolmoAct 2 (@allen_ai ): 18.9%
• Isaac GR00T N1.7 ( @nvidia): 9.5%
*Measures partial task progress, not binary completion.

[Ilya @UnderKong]

Dream Labs just raised $90M in seed funding and launched their Dream-1 API for $3/hour. 🤖💸

Where does it stack up against the rest of Physical AI?

I mapped 12 robotics foundation model providers across access, pricing, and benchmarks. 🧵👇

#Robotics #PhysicalAI #AI

[Ilya @UnderKong]

Downloadable Models & Embodied Reasoning 🌐

Robbyant: LingBot-VLA 2.0 (multi-robot hand & arm control)

@Xiaomi : Xiaomi-Robotics-0 (open weights + real-time execution)

DAMO Academy: RynnBrain 1.1 (embodied perception & 3D spatial reasoning)

[Ilya @UnderKong]

Downloadable Action Models 🛠️

@nvidia : Isaac GR00T N1.7 (weights + adaptation tools)

@allen_ai : MolmoAct 2 & 2-Think (open weights, data & code via LeRobot)

@huggingface : SmolVLA (compact 450M model for low-cost hardware)

[Ilya @UnderKong]

Hosted APIs & Research Access (cont.) 🔬

@GeneralistAI : GEN-1.5, learns physical skills from short demos (waitlist)

@SkildAI : S1 / Skild Brain for manipulation from video demonstrations

@MSFTResearch : Rho-alpha, combining vision, language, and touch

[Ilya @UnderKong]

Hosted APIs & Limited-Access Programs ☁️

Dream Labs AI : Dream-1 for two-arm manipulation ($3/h, Flash at $2/h)

@GoogleDeepMind : Gemini Robotics ER 2 for spatial reasoning & orchestration

@physical_int : π API beta for fine-tuning & hosted inference