Open a tab and you're in. Drag a robot into a scene, teleoperate it, capture training data, and export to the pipeline you already use. Nothing to set up. No GPU.
Works with
The gap between "I want to train a robot" and "I'm training a robot" used to take a week. SimArena closes that gap to a few minutes.
Before: Clone a repo. Install ROS. Fight CUDA versions. Convert URDF. Build the scene. Run out of VRAM. Start over. Alpha teams spent 5–7 days getting a single scene ready.
SimArena: Pick a scene from the gallery. Live in the editor in under two minutes. Python hooks built in. You write the task, not the scaffolding.
Before: Isaac Sim requires a workstation. MuJoCo requires a local install. Researchers without a lab GPU can't prototype.
SimArena: Physics runs in WebAssembly: MuJoCo WASM, Rapier, or PhysX. No GPU. No install. Runs on any machine.
Before: Clone a repo. Install ROS. Fight CUDA versions. Convert URDF. Build the scene. Run out of VRAM. Start over. Alpha teams spent 5–7 days getting a single scene ready.
SimArena: Pick a scene from the gallery. Live in the editor in under two minutes. Python hooks built in. You write the task, not the scaffolding.
Before: Isaac Sim requires a workstation. MuJoCo requires a local install. Researchers without a lab GPU can't prototype.
SimArena: Physics runs in WebAssembly: MuJoCo WASM, Rapier, or PhysX. No GPU. No install. Runs on any machine.
Before: Built for MuJoCo. Collaborator works in Isaac Sim. Porting takes days and always breaks the collision geometry.
SimArena: One scene exports to MuJoCo XML, MJLab, or Isaac Sim. The same robot definition runs in any engine or deploys to hardware.
Before: Teleoperate on hardware or a separate sim. Write a custom data collection script. Figure out the format. Debug the mismatch.
SimArena: Load a scene, take control, hit record. Export directly in LeRobot format. One click from scene to dataset.
Before: Nobody has automated regression testing for robot policies. Bugs surface on hardware, expensive and slow.
SimArena: Headless API connects to GitHub Actions. Deterministic execution: same seed, same result. Batch parallel. This is GitHub Actions for robot policies.
Pick a scene from the gallery or build from scratch. Place objects, configure physics, write episode behavior in Python. Success conditions, failure detection, randomization, all in the tab. No local IDE.
Physics runs locally in WebAssembly: MuJoCo WASM, Rapier, or PhysX. The control loop closes in the browser. No server round-trip. No GPU.
Hit record. Every episode saves joint states and sensor readings. Export in LeRobot format when ready. Load scene → run episodes → export data.
Export to MuJoCo XML, MJLab, or Isaac Sim. Training datasets export in LeRobot format. Switch engines from a dropdown without rewriting your scene.
Send datasets to training from the dashboard. Configure a recipe. Training runs on Fabric's serverless GPU mesh. No cluster to configure.
| Parameter | SimArena | Isaac Sim | MuJoCo Studio | Genesis | PyBullet |
|---|---|---|---|---|---|
| Boot time | Instant | Minutes | Minutes | Minutes | Seconds |
| Install required | No | Yes | Yes | Yes | Yes |
| Browser-native | Yes | No | No | No | No |
| GPU required | No | Yes | No | Recommended | No |
| Multiple physics engines | Yes (3 engines) | No | No | No | No |
| In-browser teleop + capture | Yes | No | No | No | No |
| LeRobot export | Yes | No | No | No | No |
| MuJoCo XML export | Yes | Yes | Native | — | No |
| Cloud save | Yes | No | No | No | No |
| In-browser Python editor | Yes | No | No | No | No |
| CI/CD headless API | Roadmap | No | No | No | Partial |
| Open source | Yes (MIT) | No | No | Yes | Yes |
"Environment setup that used to take weeks now takes days. Engineers open a tab and start simulating—no GPU drivers, no config fights. We're running more iterations in a month than we used to in a quarter. This is exactly what we wish we'd had when we were starting out."Early alpha teams
Yes, for in-browser simulation. Physics runs via WebAssembly in your browser tab. Cloud simulation on GPU is available on paid plans.
MuJoCo WASM closely matches native MuJoCo. For scene design and data capture, it's more than sufficient. For final training, export to a cloud engine.
URDF is the primary format. The prebuilt gallery covers common arm, mobile, and manipulation robots. A URDF builder is on the roadmap.
Training datasets in LeRobot format. Scene definitions to MuJoCo XML, MJLab, and Isaac Sim. Switch target engines from a dropdown.
Yes. The robot definition is hardware-abstracted the same URDF that runs in simulation deploys to physical hardware.
SimArena is open-source (MIT License). The in-browser simulation tier is free with 500 credits/month. Cloud features start at $99/month.
The headless API trigger from GitHub Actions is on the roadmap. Deterministic execution and batch parallel runs make automated testing practical.
URDF import is supported in V1. Full user-defined assets are on the roadmap. Agent mode generating assets from text or photos is also coming.
