The human layer of Embodied AI
Scale AI for humanoids
Human-robot interaction data for humanoid robots, generated in simulation.
Humans interact. Robots learn. We build the data layer. Users do it for free, and every session becomes training data for the teams building humanoids.
Backed by
The problem
Humanoid progress is gated by interactive human-robot data, not model size or compute.
There aren't enough physical humanoids in the world to collect the data needed to train them. So Vision-Language-Action (VLA) models starve for the diverse, interactive, multimodal data they depend on, because it barely exists yet.
How it works
We generate the decision-rich data teleop and video can't.
1.Interact
People interact with robots in simple simulations.
2.Capture
We record it all: motion, voice, language, decisions.
3.Structure
Each session becomes clean, labeled training data.
4.Train
Robot models learn from it, and we measure the gains.
5.Deploy
The skills transfer to real and simulated humanoids.
The moat
15× cheaper data that compounds while hardware burns cash.
Cost per usable data-hour
$135
Teleoperation
$9
Synthium
≈ 15× lower
Users generate the data for free because interacting with humanoids is genuinely compelling.
No teleop wages, near-zero marginal cost, and every new user makes the dataset stronger. The gap compounds into a data moat.
Proven on real hardware
Signed collaborations, live humanoids, real-world validation.
Collaboration on real humanoid hardware.
Research Collaboration Agreement (RCA).
Engineering support via Google for Startups.
On the ground in Boston.
Our edge
We capture the signal everyone else skips.
Active, not passive
The field scrapes video and teleop snapshots. We run a live loop of humans and robots interacting in real time.
Human + synthetic
Synthetic worlds need real human behavior. We're native to simulation and add what it can't fabricate: how people actually move, react, and decide.
Interaction, not dexterity
Billions chase robot hands. We own how humanoids behave safely around people, where they actually have to work.
FAQ
Frequently asked questions about Synthium.
What is Synthium?
Synthium is Scale AI for humanoids. It generates human-robot interaction data for humanoid robots in simulation. People interact with robots in live simulations, and every session becomes decision-rich, multimodal training data (motion, voice, language, decisions) that teleoperation and internet video cannot capture. Synthium is the human layer of Embodied AI.
How does Synthium work?
Users interact with robots in simple simulations. Synthium captures everything (motion, voice, language, decisions), structures each session into clean labeled training data, trains robot models on it, and the learned skills transfer to real and simulated humanoids.
Who founded Synthium?
Synthium was founded by Nicolas Duval (Co-Founder & CEO) and Nicolas Savva (Co-Founder & CTO), who met at Autodesk on the Core Rendering Team working on physically-based, photorealistic rendering.
Why is Synthium's data cheaper than teleoperation?
Synthium data costs about $9 per usable data-hour versus roughly $135 for teleoperation, about 15x lower. Users generate the data for free because interacting with humanoids is genuinely compelling, so marginal cost is near zero and every new user makes the dataset stronger.
Who is Synthium for?
Synthium is for two groups: users who want to jump into a session for free and generate training data for humanoids, and robotics teams building Vision-Language-Action (VLA) models who need interactive, multimodal datasets their models are starving for.
The human know-how inside every robot.
We start as the dataset others can't make.
Jump into a session, or partner with us on data.
Interact with a humanoid
Jump in for free. Every session generates training data for humanoids.
Start InteractingTrain your robots
Building VLAs? Get the interactive, multimodal datasets your models are starving for.