Robot-policy evaluation and simulation platform using Atom 1 plus conventional simulators to generate and grade large batches of rollouts.
Product sources: Moving Atoms official site + Y Combinator S26 company page
Moving Atoms now centers its product on RobotGym, simulation infrastructure for physical AI. Teams can build a simulation-ready digital twin from photos, video, drawings and 3D assets, run robot policies against MuJoCo physics, define evaluation budgets and retain frames, observations, commands, trajectories and grades for every run. 1
AiToolMap did not build a twin, run a robot policy, reproduce trial outcomes, measure simulator fidelity or compare RobotGym verdicts with real hardware. Y Combinator corroborates the company and earlier Atom world-model direction but is not an independent product benchmark. No eligible fixed-panel numeric evidence was found. The 8.4 rating is LOW-confidence and not rank eligible.
Simulation should reduce risky physical experimentation, not authorize deployment by itself. Robotics teams should define conservative success criteria, retain replayable evidence, validate edge cases on hardware, preserve emergency-stop and workspace safety controls, and treat simulator pass/fail as one layer in a broader physical validation process.
Transparency is the standout feature. RobotGym is built around inspectable evidence rather than a single model-generated success label: runs preserve exact observations, commands and trajectories, can be replayed and compared, and public environments expose concrete examples. This makes the product useful not only for training but for evaluation and debugging. 1 2
Moving Atoms is also unusually explicit about its own experimental limits. The Atom history page reports a one-call agent succeeding in 3 of 6 development cases, a later program pilot at 4 of 4 while noting that two delegated to a now-unavailable hosted policy, and six arms across forty LIBERO tasks passing startup/observation checks while explicitly saying that this was infrastructure validation, not policy success. That level of qualification materially improves trust. 3
Pricing is now public and materially more accessible than the old design-partner-only record: Research is $100/month, Lab $300/month and Enterprise by arrangement. Verified new accounts receive $100 in welcome credits with no card required. 1 4
The core unanswered question is sim-to-real validity. A simulation platform can make development faster and more reproducible, but success in a twin does not automatically predict robust physical-world behavior under calibration error, sensor noise, contact uncertainty and hardware variation. AiToolMap found no independent benchmark establishing that predictive relationship for RobotGym.
Current scope is also narrow. The Atom page says the present workspace exposes the Panda reference robot only. Public trials are useful evidence of engineering maturity, but they remain first-party and should not be interpreted as general robot-policy success rates. 3
Products matched by shared tasks, audience, workflow, product type and capabilities.