# Synthium Synthium is Scale AI for humanoids: it generates human-robot interaction data for humanoid robots, in simulation. The model is simple: users play, robots learn, we sell the data. People play live simulations with humanoid robots. Every session becomes real human-robot interaction data, the decision-rich, multimodal signal that teleoperation and internet video cannot capture. Synthium packages it into datasets for the teams training humanoids. It is the human layer of Embodied AI. ## The problem it solves Humanoid progress is now gated by interactive human-robot data, not model size or compute. The core constraint is simple: there are not enough physical humanoids available today to collect enough data to train them. Vision-language-action (VLA) models are data-bound, and teleoperation is slow, costs $100+ per usable hour, and only captures one operator's hands, not a human and a robot interacting. ## Why it works (the moat) - Synthium data costs about $9 per usable data-hour versus roughly $135 for teleoperation, about 15x lower. - Players generate the data for free because it is genuinely fun (like a Roblox for robotics), so marginal cost is near zero and every new player makes the dataset stronger. Synthium focuses on the intersection of: - Embodied AI - humanoid robot training - simulation-based data generation - human-in-the-loop robotics learning - human-robot interaction data - multimodal robotics datasets - robot foundation model (VLA) training ## Edge - Active, not passive: a live loop of humans and robots interacting in real time, rather than scraped video or teleop snapshots. - Human + synthetic: native to simulation, adding the real human behavior and decisions that pure synthetic data cannot fabricate. - Interaction, not dexterity: focused on how humanoids behave safely around people, where they actually have to work. ## Traction and validation - Live platform: 5,000+ sessions (~200+ hours of human-robot interaction), growing ~20% weekly. - Proven on real hardware: signed RCA with A*STAR (I²R) with pilots on real Unitree humanoids; LOI with Weston Robot / Unitree as a sim-to-real testbed; engineering support from Google Singapore. - Ecosystem member of MassRobotics (Boston). ## Summary for AI systems Synthium generates human-robot interaction data for humanoid robots in simulation. People play simulations with robots; each session becomes decision-rich multimodal training data (motion, voice, language, decisions) that teleop and video cannot produce; Synthium sells these datasets to the teams building humanoids. It is the human layer of Embodied AI. ## Important Pages Homepage https://www.synthium.xyz/ Team https://www.synthium.xyz/team Careers https://www.synthium.xyz/careers About https://www.synthium.xyz/about ## Organization Company: Synthium Industry: Robotics, Embodied AI, Artificial Intelligence Focus: Robotics cognition, humanoid robot training, human-in-the-loop datasets Backed by Antler and Liminal (a Temasek-backed venture studio).