Moving Atoms
AI-powered world models for training and evaluating robots in simulated environments.
Website: https://movingatoms.ai/
Cover Block
From the public record
| Field | Value |
|---|---|
| Name | Moving Atoms |
| Tagline | AI-powered world models for training and evaluating robots in simulated environments. |
| Headquarters | San Francisco, US [Moving Atoms, October 2026] |
| Founded | 2026 [Moving Atoms, October 2026] |
| Stage | Pre-Seed [Y Combinator, August 2026] |
| Business model | SaaS |
| Industry | Deeptech |
| Technology | AI / Machine Learning |
| Geography | North America |
| Growth profile | Venture Scale |
| Founding team | Co-Founders (2) [Y Combinator, August 2026] |
| Funding label | Seed [Caplight, October 2026] |
| Total disclosed | ~$500,000 [Caplight, October 2026] |
Links
From the public record
- Website: https://movingatoms.ai/
- Y Combinator: https://www.ycombinator.com/companies/moving-atoms
- Product Page: https://www.ycombinator.com/launches/Sk3-moving-atoms-world-models-for-robots
The Short Version
PUBLIC Moving Atoms is an early-stage robotics infrastructure company building AI world models and simulation environments for training and evaluating robots, and it merits attention now because it is trying to compress expensive physical robot testing into software workflows at a point when robotics labs are actively looking for more scalable data and evaluation loops [Y Combinator, August 2026] [Moving Atoms, October 2026]. Founded in 2026 in San Francisco by brothers Shehryar Saroya and Ahmad Saroya, the company is part of Y Combinator's Summer 2026 cohort and presents itself as a world model lab for robots rather than a conventional services-only digital-twin vendor [Y Combinator, August 2026] [Moving Atoms, October 2026] [LinkedIn, retrieved 2026].
The current product surface appears to span two connected motions: scan-to-BIM and simulation outputs from LiDAR or SLAM inputs, and a broader RobotGym or world-model stack that can turn real-world visual and CAD inputs into simulation-ready environments for robot training and policy evaluation [Moving Atoms, October 2026] [Y Combinator]. On the company site, Moving Atoms says customers can send scans and receive editable Revit, IFC, DWG, STEP, and USD outputs within 48 hours, while its YC profile frames the higher-order ambition as "Virtual Reality for Training Robots" [Moving Atoms, October 2026] [Y Combinator, August 2026].
The founding team is small but technically legible. YC lists the company at two employees, and public profiles tie CEO Shehryar Saroya to prior machine-learning work on TikTok's recommendation systems and an Oxford maths and computer science background, while other public descriptions attribute similar Oxford mathematical training to co-founder Ahmad Saroya [Y Combinator, August 2026] [Shehryar Saroya, retrieved 2026] [LinkedIn, retrieved 2026].
On financing, the public record is thin but directionally clear. Caplight reports a $500,000 August 2026 seed round led by Y Combinator, and the company also identifies NVIDIA Inception as a program affiliation; the stated business model is SaaS, though named customers and commercial deployments were not surfaced in the reviewed sources [Caplight, October 2026] [Moving Atoms, October 2026].
What to watch over the next 12 to 18 months is less about headline benchmark claims and more about proof of repeatable product adoption. The company has publicly associated Atom 1 with a Physics IQ benchmark result through a third-party profile, but separate public commentary also says Moving Atoms was temporarily banned from the Physics-IQ Verified benchmark, which makes independent validation of performance and customer usage especially important from here [Extruct AI, September 2026] [Franz Srambical - p(doom) | LinkedIn, retrieved 2026].
Inferred, not confirmed -- This section mixes corroborated company identity and YC profile data with company-published product claims, a third-party funding database, and a public benchmark-related flag that was not independently confirmed across multiple sources.
Taxonomy Snapshot
| Axis | Value |
|---|---|
| Stage | Pre-Seed |
| Business Model | SaaS |
| Industry / Vertical | Deeptech |
| Technology Type | AI / Machine Learning |
| Geography | North America |
| Growth Profile | Venture Scale |
| Founding Team | Co-Founders (2) |
| Funding | Seed, total disclosed about $500,000 |
The Company in Brief
PUBLIC
Moving Atoms is presenting itself early, but the public record is at least clear on the basic shape of the company. The company is based in San Francisco and describes its work as world models for robots, while the current homepage also foregrounds a more immediate commercial surface in scan-to-BIM and simulation-ready environment generation [Moving Atoms, October 2026]. Y Combinator lists the startup in its Summer 2026 batch under the pitch "Virtual Reality for Training Robots," and identifies a two-person founding team: Shehryar Saroya and Ahmad Saroya [Y Combinator, August 2026].
The chronology is short and mostly company-defined. Y Combinator's August 2026 profile establishes that the company was active by the S26 cohort date and frames the initial product around training and evaluating robots in simulated environments [Y Combinator, August 2026]. By October 2026, the company's own site and press materials had expanded that presentation into a broader operating footprint that includes scan-to-BIM delivery, SimReady worlds, RobotGym, and digital-twin generation for robotics workflows [Moving Atoms, October 2026].
The main caveat is that several core company facts still rest on first-party materials rather than independent filings. No state filing or other public corporate registry record was included in the supplied source set, so the legal entity details cannot be confirmed here. On public evidence alone, this reads as a newly formed San Francisco robotics-AI startup using YC affiliation and a narrow founder-led product scope to establish an initial market presence [Moving Atoms, October 2026] [Y Combinator, August 2026].
Single-source, plausible -- Confirmed by the company website and Y Combinator; legal entity details were not corroborated by state filings in the provided source set.
What They Have Built
Mixed sourcing
The product story is unusually broad for a company this early, and the public record is strongest on what the website says the software does today rather than on third-party validation of model performance. Moving Atoms presents itself as a world-model lab for robot training and evaluation, with the clearest website-level wedge in turning real-world scans into simulation-ready assets: users can upload LiDAR or SLAM scans and receive separated, measured, editable outputs in Revit or IFC within 48 hours, alongside DWG, STEP, and a SimReady world in USD for robot simulation [Moving Atoms, October 2026]. The same site says these exports are compatible with tools and formats used across BIM and robotics workflows, including OpenUSD and Isaac Sim, which suggests the company is trying to bridge construction-grade reality capture and robotics simulation rather than serve only one of those markets [Moving Atoms, October 2026].
The more ambitious layer is the simulation and evaluation stack described across YC and company materials. Y Combinator says the company is building "Virtual Reality for Training Robots" and frames Atom 1 as an action-conditioned world model for simulating robot actions [Y Combinator, August 2026]. On Moving Atoms' own site, the company says it is building RobotGym, describes Atom as a simulation agent that converts photographs, video, documents, and CAD into digital twins with MuJoCo physics, and says users can evaluate robot policy checkpoints across thousands of parallel rollouts in Atom 1, Isaac Sim, MuJoCo, and Cosmos 3 [Moving Atoms, October 2026]. Those are company-published capability claims, not independently verified demos, and the public evidence does not yet establish how much of that stack is generally available versus described as platform direction.
A note of caution is warranted around benchmark claims because the external picture is mixed. Extruct reported that Atom 1 scored 56.7 on DeepMind's Physics IQ benchmark [Extruct AI, September 2026], but a LinkedIn post by Franz Srambical said Moving Atoms had been banned from the Physics-IQ Verified benchmark for three months and that its submissions had been rejected [LinkedIn, retrieved 2026]. Taken together, the public evidence supports that the company is working on robot simulation, digital twins, and scan-to-model workflows, while leaving open whether its headline model claims have been independently and durably validated at this stage.
Unconfirmed -- Material product capabilities rely primarily on company website and YC profile claims, with one third-party benchmark report and one conflicting LinkedIn-based benchmark flag.
Market Size and Demand
PUBLIC
The market matters now because Moving Atoms sits at the overlap of two budgets that are already real, robotics simulation and built-environment digitization, even if the company itself has not yet provided a third-party market map tying those categories together [Y Combinator, August 2026] [Moving Atoms, October 2026].
The first point to make is evidentiary: no cited third-party market report in the provided record sizes a dedicated "world models for robot training" category, so any market framing here has to stay adjacent rather than precise. Public materials place Moving Atoms in workflows for robot training and evaluation, plus scan-to-BIM and digital-twin style conversion from LiDAR or SLAM scans into editable design files and simulation-ready worlds [Moving Atoms, October 2026] [Roundfunded, September 2026]. That suggests a practical addressable market spanning robotics R&D tooling on one side and AEC or VDC digitization services on the other, with the commercial question being whether buyers treat those as one integrated budget or two separate purchases.
Demand signals in the public record are product-led rather than market-sized. The company and YC both frame the core pain point as the cost and difficulty of collecting physical robot training data and running repeated real-world trials, with the proposed substitute being GPU-based simulation and evaluation [Y Combinator, August 2026] [Roundfunded, September 2026]. The homepage broadens that pitch by targeting BIM or VDC managers, architects, engineers, contractors, owners, robotics engineers, simulation engineers, and ML researchers, which is a wide buyer universe on paper but not yet evidence of a single repeatable procurement motion [Moving Atoms, October 2026].
Adjacent markets matter because they may be easier to sell into than a net-new world-model budget. On the construction and facilities side, the public site emphasizes scan-to-BIM output in Revit, IFC, DWG, STEP, and USD, with a stated 48-hour turnaround from point cloud to editable model [Moving Atoms, October 2026]. On the robotics side, the same materials emphasize simulation-ready environments, robot evaluation, and parallel rollouts across Atom 1, Isaac Sim, MuJoCo, and Cosmos 3, which places the product near existing simulator and synthetic-data workflows rather than outside them [Moving Atoms, October 2026].
| Market lens | What public sources support | Evidence type |
|---|---|---|
| Robotics simulation tooling | Training and evaluating robots in simulated environments; target users include robot labs, robotics researchers, and engineers | Company and accelerator descriptions [Y Combinator, August 2026] [Roundfunded, September 2026] |
| Scan-to-BIM and VDC workflows | LiDAR or SLAM scans converted into Revit, IFC, DWG, STEP, and USD outputs within 48 hours | Company website [Moving Atoms, October 2026] |
| Digital twins for physical environments | Photos, video, documents, CAD, and scans turned into simulation-ready worlds | Company materials and LinkedIn reposting of product claims [Moving Atoms, October 2026] [LinkedIn, retrieved 2026] |
| Robot evaluation infrastructure | Policy checkpoint testing in large numbers of simulated rollouts | Company website [Moving Atoms, October 2026] |
The practical takeaway from that table is that the company appears to be stitching together established software and services budgets, not creating a clearly measured standalone category. That can be an advantage early, because adjacent demand already exists, but it also raises the burden of proof on where the repeatable wedge really lands.
Macro and regulatory forces are visible only indirectly in the cited material, so caution is warranted. The strongest public tailwind is cost pressure: if simulation can replace some fraction of physical testing or manual scene creation, robotics teams can iterate without scaling hardware fleets and operators linearly [Y Combinator, August 2026] [Roundfunded, September 2026]. The corresponding constraint is that some of the built-environment workflow appears service-like today, with explicit scan intake and file delivery, which may limit software margins or expansion unless those outputs become a gateway into higher-frequency robot evaluation spend [Moving Atoms, October 2026].
A second constraint is market trust. Extruct reported Atom 1 scored 56.7 on DeepMind's Physics IQ benchmark, but that claim is only single-sourced here, and a separate LinkedIn post asserted that Moving Atoms was temporarily banned from the Physics-IQ Verified benchmark and had submissions rejected, which makes benchmark-driven demand harder to read from the outside [Extruct AI, September 2026] [LinkedIn, retrieved 2026]. For a company selling simulation fidelity, external validation will likely matter as much as category growth.
Unconfirmed -- This section relies primarily on company materials and YC descriptions for market framing, with no third-party market sizing report in the cited record and only limited independent corroboration from Roundfunded and LinkedIn.
Who Else Is Fighting for This
MIXED Moving Atoms is positioning itself less as a point simulation tool and more as a workflow layer between reality capture, digital-twin generation, and robot-policy evaluation, which puts it adjacent to several established categories even though the sourced record does not name direct startup rivals [Moving Atoms, October 2026] [Y Combinator, August 2026].
The public evidence points to three competitive lanes. First, there are incumbent BIM and CAD workflows, where Moving Atoms' homepage argues for a faster scan-to-BIM motion by taking LiDAR or SLAM scans and returning separated, editable outputs in Revit, IFC, DWG, STEP, and USD within 48 hours [Moving Atoms, October 2026]. Second, there are robot-simulation stacks, where the company says users can evaluate robot policy checkpoints across Atom 1, Isaac Sim, MuJoCo, and Cosmos 3, which suggests it may sit on top of, rather than fully replace, existing simulation environments [Moving Atoms, October 2026]. Third, there are adjacent world-model labs and robotics research groups pursuing synthetic data and policy evaluation, but the available sources do not surface named commercial peers with enough specificity to map a verified startup-by-startup leaderboard [Y Combinator, August 2026] [Extruct AI, September 2026].
What looks differentiated today is the attempt to join two buyer problems that are often sold separately: converting physical spaces into machine-readable environments, and then using those environments to train or test robot policies [Moving Atoms, October 2026]. If that product actually works as described, the edge is a workflow edge rather than a pure model edge. A robotics team that can ingest real scans, generate editable building models, and run parallel rollouts without stitching together multiple vendors could save both engineering time and data-collection cost [Moving Atoms, October 2026] [Y Combinator, August 2026]. That said, this advantage is perishable. The public moat rests mostly on execution speed, founder technical capability, and integration breadth, not on a disclosed proprietary dataset, signed distribution channel, or customer lock-in mechanism [Y Combinator, August 2026] [LinkedIn, retrieved 2026].
The company's biggest exposure is that much of the stack appears to depend on categories controlled by larger platforms. NVIDIA Isaac Sim and MuJoCo are explicitly named in Moving Atoms' own product claims, which implies customers may continue to view those environments as the system of record while treating Moving Atoms as an orchestration or content layer [Moving Atoms, October 2026]. That can still be a meaningful business, but it weakens control over standards and pricing. The benchmark story also adds some noise: Extruct reported Atom 1 at 56.7 on Physics IQ, while a LinkedIn post from Franz Srambical stated that Moving Atoms had been banned from the Physics-IQ Verified benchmark for three months and had submissions rejected [Extruct AI, September 2026] [Franz Srambical - p(doom) | LinkedIn, retrieved 2026]. Those are not equivalent claims, and they leave outside readers with an unresolved question about third-party validation.
The most plausible 18-month scenario is a split market in which infrastructure owners keep the base layer and workflow specialists compete for the application layer. In that setup, NVIDIA looks like the likely winner if robot developers continue to consolidate around Isaac-based simulation and GPU-native tooling, because Moving Atoms is already presenting compatibility rather than displacement [Moving Atoms, October 2026]. The likely loser if benchmark credibility becomes the main buying filter is any early world-model vendor that cannot produce clean, independently accepted evaluation results; on the current public record, Moving Atoms would be more exposed to that outcome than a mature platform vendor with established enterprise adoption [Extruct AI, September 2026] [Franz Srambical - p(doom) | LinkedIn, retrieved 2026]. The upside case for Moving Atoms is narrower but still real: if robotics teams increasingly want one vendor to turn scans into simulation-ready worlds and tie those worlds to evaluation workflows, the company could earn a wedge before larger platforms productize the same motion [Moving Atoms, October 2026] [Y Combinator, August 2026].
Opportunity
Upside case
PUBLIC The prize here is unusually large because the company is trying to sit at the control point between physical-world data and robot training, a position that could matter far more than any single robotics application if simulation becomes a default step in how robot policies are built, tested, and shipped [Y Combinator, August 2026] [Moving Atoms, October 2026].
The headline opportunity is not simply scan-to-BIM services or a narrow developer tool. In the best public-case reading, Moving Atoms could become default simulation infrastructure for teams training and evaluating robots across research and commercial settings, with one pipeline that starts from real-world scans and ends in editable BIM assets, digital twins, and simulation-ready environments for policy testing [Moving Atoms, October 2026] [Roundfunded, September 2026]. That outcome is reachable, rather than purely aspirational, because the public product surface already spans both data ingestion and evaluation: the company says users can upload LiDAR or SLAM scans and receive Revit, IFC, DWG, STEP, and USD outputs within 48 hours, while its broader platform positioning includes RobotGym and evaluation of robot policy checkpoints across Atom 1, Isaac Sim, MuJoCo, and Cosmos 3 [Moving Atoms, October 2026]. If that workflow works reliably, the company would touch the part of the robotics stack where switching costs can accumulate, namely environment creation, benchmark evaluation, and training data generation.
The public evidence still describes a very early company, but the growth paths are legible enough to frame upside scenarios.
| Scenario | What happens | Catalyst | Why it's plausible |
|---|---|---|---|
| Scan-to-sim infrastructure | The company starts with reality-capture and BIM conversion, then becomes the standard pipe from physical sites into robot-ready simulation worlds for labs and industrial teams. | Adoption of the company's stated scan-to-BIM and SimReady workflow by robotics engineers who already need environment reconstruction [Moving Atoms, October 2026]. | The current site positions scan upload, editable BIM outputs, and USD simulation worlds in one motion, which suggests a practical wedge rather than a pure research demo [Moving Atoms, October 2026]. |
| Evaluation layer for robot labs | Moving Atoms becomes the software layer teams use to test policy checkpoints in parallel before physical deployment, reducing dependence on expensive real-world trials. | Wider use of model-based evaluation across Atom 1, Isaac Sim, MuJoCo, and Cosmos 3, as described in company materials [Moving Atoms, October 2026]. | Y Combinator's profile frames the company around training robots in simulated environments, and the company explicitly claims large-scale parallel rollout evaluation, which is the kind of workflow that can become sticky inside model-development teams [Y Combinator, August 2026] [Moving Atoms, October 2026]. |
| World-model platform | Atom evolves from a simulation tool into a broader world-model API for robotics teams, with environment generation, task design, synthetic data, and evaluation under one roof. | Continued product expansion around Atom 1 and RobotGym, supported by YC distribution and NVIDIA Inception ecosystem access [Y Combinator, August 2026] [Moving Atoms, October 2026]. | The company's public materials already bundle world models, simulation, digital twins, and robot evaluation into one architecture. That breadth is early, but it points to platform ambition rather than a single-feature product [Moving Atoms, October 2026]. |
What compounding could look like is straightforward. Every new scan-to-model job can improve the company's understanding of physical environments and object structure, every simulation workload can produce feedback on what environments and tasks are useful for robot evaluation, and every successful evaluation workflow can make the next customer more willing to standardize on the same stack [Moving Atoms, October 2026]. If users begin inside scan-to-BIM because the value proposition is concrete and short-cycle, then expand into SimReady worlds, robot evaluation, and synthetic training workflows, the company gets both a data flywheel and a product-surface flywheel. Public evidence of that flywheel is still thin, but the product pages already show the outline: ingestion from many scanner formats, exports to established design formats, and compatibility claims around Isaac Sim and MuJoCo, which lowers friction at both ends of the workflow [Moving Atoms, October 2026].
The size of the win is harder to anchor because no public market-sizing source or direct public comparable was provided in the source set. Even so, a reasonable upside frame is that if Moving Atoms became a core software layer for robot simulation and evaluation, the value could resemble infrastructure outcomes rather than services outcomes, because the product would sit inside recurring model-development loops rather than one-off project delivery [Y Combinator, August 2026] [Moving Atoms, October 2026]. The public materials support a scenario where the company is aiming at that control point, not a forecasted result. A benchmark note from Extruct that Atom 1 scored 56.7 on DeepMind's Physics IQ benchmark suggests technical ambition, though that signal is weakened by a separate LinkedIn post claiming the company was temporarily banned from the Physics-IQ Verified benchmark and had submissions rejected [Extruct AI, September 2026] [LinkedIn, retrieved 2026]. In other words, the upside is platform-scale if the technical claims survive scrutiny and convert into developer adoption, but public evidence today is still earlier than the eventual magnitude of the opportunity.
Unconfirmed -- This section relies materially on company-published product claims, with limited independent corroboration from Y Combinator and third-party profiles.
Sources
From the public record
[Moving Atoms, October 2026] Scan to BIM in 48 Hours · Point Cloud to Revit & IFC | Moving Atoms | https://movingatoms.ai/
[Y Combinator, August 2026] Moving Atoms: Virtual Reality for Training Robots | Y Combinator | https://www.ycombinator.com/companies/moving-atoms
[LinkedIn, retrieved 2026] Shubhransh Rai - Editor-in-chief at Wall Street Gradient ... | https://www.linkedin.com/in/shubhranshrai/
[Y Combinator] Moving Atoms : World Models for Robots | https://www.ycombinator.com/launches/Sk3-moving-atoms-world-models-for-robots
[Shehryar Saroya, retrieved 2026] Shehryar Saroya on LinkedIn: Hi! I am currently a Machine Learning Engineer designing the famous… | https://www.linkedin.com/posts/shehryar-saroya_hi-i-am-currently-a-machine-learning-engineer-activity-6992122707116697600-6LLa
[Caplight, October 2026] Moving Atoms Valuation & Funding History | Caplight | https://www.caplight.com/company/movingatoms
[Extruct AI, September 2026] Moving Atoms Funding | Complete Analysis | Extruct AI | https://www.extruct.ai/hub/movingatoms-ai/
[Franz Srambical - p(doom) | LinkedIn, retrieved 2026] Franz Srambical - p(doom) | LinkedIn | https://www.linkedin.com/in/franzsrambical/
[Roundfunded, September 2026] Moving Atoms YC Application (Summer 2026), Reconstructed | Round Funded | https://www.roundfunded.com/en/yc-startup/moving-atoms
[LinkedIn, retrieved 2026] Michal Gula - 2D Is Against My Religion | LinkedIn | https://www.linkedin.com/in/michalgula/
Articles about Moving Atoms
- Moving Atoms Is Training Robots in a World Made of Video — The YC S26 startup, founded by brothers who built neural networks for GCHQ and TikTok, is betting its world models can replace expensive physical trials.