Synthium was founded by Nicolas Duval (CEO) and Nicolas Savva (CTO), two repeat founders who met their current problem from opposite technical angles. Duval spent two decades in software architecture and AI, including leadership roles at Autodesk and Adobe, where he helped shape Adobe Express across cloud infrastructure, CAD, and generative AI. Savva holds an MS/PhD in computer graphics and computer vision from Cornell, where he lectured and researched neural image captioning and photorealistic rendering, and holds an O-1 visa for individuals of extraordinary ability. Together they're building what they call "the human layer of Embodied AI."

Data, not dexterity, is the bottleneck

The company's core thesis is that humanoid progress isn't gated by model size or compute — it's gated by a simple physical shortage: there aren't enough real humanoid robots in the world to generate the interactive, multimodal data that Vision-Language-Action (VLA) models need to learn from. Synthium's answer is to let people play with robots in simulation instead. Every session — motion, voice, language, decisions — gets captured, structured into clean labeled training data, and used to train robot models, with the resulting skills transferring to both real and simulated humanoids. It's a five-step loop the company describes simply as play, capture, structure, train, deploy.

A cost moat built on fun, not wages

Where traditional teleoperation costs roughly $135 per usable data-hour, Synthium says its model brings that down to around $9 — roughly 15x cheaper — because players generate the data for free, drawn in by gameplay the company compares to "a Roblox for robotics." With no teleop wages and near-zero marginal cost, every new player strengthens the dataset rather than adding to the bill, compounding into a data moat that widens as more people play. The company also frames its edge as being about what it captures rather than just cost: an active, real-time loop of humans and robots interacting, rather than scraped video or static teleop snapshots — and a focus on how humanoids behave safely around people, rather than the dexterity problem most of the field is chasing.

Synthium says it already has a research collaboration agreement and hands-on work with real humanoid hardware, along with engineering support through Google for Startups. The company is backed by Antler and Liminal, a Temasek-backed venture studio, and is currently building out both its consumer play-to-generate-data product and a data offering for teams training their own VLA models.

A network neither founder knew existed

On the value of Embark, the team is blunt: "The doors you have opened for us, I didn't even know existed in SF. Worked really really well for us. The program was very well structured…I am more worried about how would you do this again with such flawless execution."

‍