Moving Atoms
Moving Atoms provides simulation infrastructure for physical AI, enabling robots to learn and deploy at scale through SimReady assets, simulated training data, and evaluation platforms. Its flagship product, RobotGym, lets users build digital twins from a single photograph and run robot training and evaluation scenarios in the browser. The platform supports editable scenes with real physics, declared evaluation budgets, and downloadable run records for transparent performance assessment.
- Artificial Intelligence
- Industrial Automation
- Robotics
- Software Only
Funding
Founders
Product
Problem
Training and evaluating physical AI systems requires large volumes of high-quality, physically accurate simulation data, but building these environments is time-consuming and expensive. Existing simulation tools often lack the visual realism and physical fidelity needed to transfer learned behaviors to real-world robots, and evaluation processes frequently lack transparency and reproducibility.
Solution
Moving Atoms provides a simulation infrastructure platform that generates SimReady assets, simulated training data, and evaluation environments for physical AI development. Its RobotGym product enables users to reconstruct complete 3D scenes from a single photograph, with editable geometry, materials, lighting, and robot placement controlled through natural language chat commands. The platform supports modular skill creation through policy training, human teleoperation, and video-to-sim transfer, while its evaluation harness runs tasks with declared budgets, records every attempt, and provides independent verdicts with full run history. Built environments retain their editable Blender scenes, GLB files, and source images, allowing users to choose what to publish and maintain full control over their data.
Target Audience
Primary customers are industrial automation teams validating cells before commissioning, manipulation researchers testing grasps and placements, and embodied AI developers running model evaluations with transparent, reproducible benchmarks.
Features
- Scene reconstruction from a single photograph into editable 3D geometry with real physics underneath
- Chat-based scene editing that modifies geometry, materials, lighting, robot placement, and cameras without altering the underlying model
- Robot ingestion harness supporting URDF control interfaces and system identification
- Skill creation tools including modular composition, policy training, human teleoperation, and video-to-sim transfer
- Evaluation harness with declared budgets, per-run limits, and recorded termination reasons for every attempt
- Downloadable, replayable run records including model IDs, tokens per response, program steps, and settling metrics
- Built-in regression testing and validation with frozen scenes, seeds, and budgets for reproducible benchmarks
- Support for multiple frontier models including GPT-6 Astra and Claude Fable 5 with side-by-side attempt comparison