You upload a video of a cluttered workshop. A few minutes later, inside a simulation, a robotic arm is trying to pick a wrench off a workbench, its digital fingers fumbling with the physics of weight and friction. The training run costs a few dollars in cloud compute, not thousands in hardware and labor. This is the promise Moving Atoms sells: virtual reality for robots, where the world is made of learned physics, not hand-coded rules.
Founded in 2026 by brothers Shehryar and Ahmad Saroya, the San Francisco startup is building what it calls "world models",AI systems trained on internet video that can simulate realistic environments for robots to practice in [Moving Atoms, October 2026]. The bet is that GPU-based simulation can finally become cheap and accurate enough to displace a significant chunk of the grueling, expensive process of training robots in the real world.
The Algorithmic Wedge
The founders' backgrounds read like a targeted recruitment drive for this specific problem. Shehryar Saroya, the CEO, worked as a machine learning engineer on TikTok’s core video recommendation algorithm, a system he describes as making "small nudges [that] change thousands of years of human time per day" [Shehryar Saroya, retrieved 2026]. His brother and CTO, Ahmad, studied mathematics at Oxford, won a British Mathematical Olympiad gold medal, and previously built neural networks for the UK government’s intelligence agency, GCHQ [Perplexity Sonar Pro Brief].
This combination is the company's foundational insight. The TikTok experience speaks to scaling complex models that learn from vast, unstructured video data. The mathematics and government work speaks to rigorous, security-conscious engineering of those systems. Moving Atoms is applying that blend to robotics, a field historically bottlenecked by data. "Instead of buying 100 robot arms," the thinking goes, "buy 100 GPUs and simulate 10,000 arms."
A Model and a Service
The company's first public product is Atom 1, a model pretrained on internet video and fine-tuned to understand physics and robot actions [Perplexity Sonar Pro Brief]. It forms the core of their RobotGym simulation infrastructure. But the commercial entry point is more pragmatic: a Scan-to-BIM service. Customers can send LiDAR or SLAM scans from any site and, within 48 hours, receive a detailed, editable digital model with every object separated and ready for design software like Revit [Moving Atoms, October 2026].
This service isn't just a revenue line. It's a data flywheel. Each scan feeds the company's understanding of real-world geometry and clutter, making its core simulation models more accurate. It’s a classic wedge,solving an immediate, painful problem for construction and engineering teams today to build the dataset for the grander robotics ambition tomorrow.
The early technical signals are strong, if messy. In September 2026, the Atom 1 model reportedly scored 56.7 on DeepMind's Physics IQ benchmark, topping the leaderboard [Extruct, September 2026]. That same month, however, the benchmark administrators banned Moving Atoms from the platform for three months and rejected its submissions, according to a LinkedIn post from a researcher involved with the benchmark [Franz Srambical - p(doom) | LinkedIn, retrieved 2026]. The post did not specify a reason, but such bans in academic benchmarks often relate to submission violations or questions about methodology. For a startup staking its claim on superior physics understanding, it's a contradictory debut: a flash of top-tier performance followed by a very public timeout.
The Two-Person Traction Engine
For now, Moving Atoms is the two founders and a $500,000 seed round led by Y Combinator in the summer of 2026, with additional backing from NVIDIA’s Inception program [Caplight, October 2026] [Moving Atoms, October 2026]. The investor lineup is a perfect fit: Y Combinator for early-stage product rigor, and NVIDIA for the essential GPU infrastructure and ecosystem credibility in AI and simulation.
| Founder | Role | Key Background |
|---|---|---|
| Shehryar Saroya | Co-Founder & CEO | ML Engineer, TikTok core recommendation algorithm; MSc, Maths & Computer Science, Oxford [Perplexity Sonar Pro Brief][Shehryar Saroya, retrieved 2026] |
| Ahmad Saroya | Co-Founder & CTO | Built neural networks for GCHQ; British Mathematical Olympiad gold medalist; Mathematics, Oxford [Perplexity Sonar Pro Brief] |
The path from here involves navigating a few inherent tensions in the bet. The market for high-fidelity robot simulation is not new, with entrenched players built on years of proprietary physics engines. Moving Atoms is betting that a data-driven, AI-native approach learned from the real world will outperform those hand-crafted systems. Furthermore, the business model straddles a services-led motion (Scan-to-BIM) and a platform SaaS future (RobotGym subscriptions). Balancing these without becoming a consultancy is a classic startup tightrope.
The Cultural Question in the Code
The most telling detail might be in the founders' prior obsessions. One tuned an algorithm that mastered human attention. The other built systems for an organization that masters secrets. Now, they are building a machine that must master the silent, implicit rules of the physical world,gravity, texture, momentum. The product they are shaping answers a deeper question: in a race to build intelligent machines, is the final, most valuable training data simply everything we've already recorded? Moving Atoms is betting that the answer is hidden in the pixels of every mundane video online, waiting to teach a robot how to pick up a wrench.
Sources
- [Moving Atoms, October 2026] Press kit · Moving Atoms | https://movingatoms.ai/press/
- [Moving Atoms, October 2026] Scan to BIM in 48 Hours · Point Cloud to Revit & IFC | https://movingatoms.ai/
- [Perplexity Sonar Pro Brief] Moving Atoms company briefing
- [Shehryar Saroya, retrieved 2026] Shehryar Saroya professional profile on LinkedIn
- [Extruct, September 2026] Moving Atoms Funding | Complete Analysis | https://www.extruct.ai/hub/movingatoms-ai/
- [Franz Srambical - p(doom) | LinkedIn, retrieved 2026] LinkedIn post regarding Physics-IQ benchmark
- [Caplight, October 2026] Moving Atoms Valuation & Funding History | https://www.caplight.com/company/movingatoms
- [Y Combinator, August 2026] Moving Atoms: Virtual Reality for Training Robots | https://www.ycombinator.com/companies/moving-atoms