SimForge is open

Accounts are open to everyone. A scenario dataset, seven SimForge Worlds, a runtime free for non-commercial use, and an open toolkit.

SimForge simulates driving on real places. We survey a place, rebuild it as a map a simulator can drive (lanes, junctions, signals, kerbs) and render it in 3D with Helios Engine, our renderer. Recorded drives become scenarios on those maps, and a policy can take over at any moment and drive the scenario closed loop.

Until now that ran in our cloud, for accounts we invited. This research release opens it up: anyone can create an account, and the parts researchers need run on their own machine:

• Accounts for everyone: sign in and start, no waitlist.
• A scenario dataset on Hugging Face, CC BY-NC 4.0.
• Seven SimForge Worlds, real places as HD maps and digital twins, free with sign-in.
• Helios Engine and the simulator as a binary, free for non-commercial use.
• An open toolkit under Apache-2.0: the simforge CLI and Python clients.
• SimForge Gym, an RL environment with a measured, reproducible PPO baseline.

Everything is free for non-commercial use, research included. A commercial licence is available for every part.

SimForge Studio in the browser is free to explore with the same account. SimForge Cloud, the paid compute behind it (3D renders, evaluations, transfers, Recast and --cloud runs on our GPUs), is open in beta. New accounts start with $10 in free credits; credit purchases aren't open during the beta.

Accounts are open

Create an account on simforge.ai and sign in. The same account signs in the simforge CLI on your machine and unlocks the runtime and SimForge Worlds downloads, and opens SimForge Studio in the browser, free to explore. There is no waitlist and no invitation step.

The scenario dataset

A scenario is a driving situation the simulator can replay: the road, the other road users, what each of them does, and where your car starts and is going. Each scenario in the dataset holds the positions and motion of every road user on a SimForge map, with no camera images and no lidar. They include situations such as a pedestrian crossing mid-block, a car pulling out of a side street in front of you, or an unprotected left turn.

The first release holds 24,483 driving scenarios on 7 SimForge Worlds.

• huggingface.co/datasets/SIMFORGE-AI/simlog-25k
• Each scenario ships as a SimForge scenario document, a trace, and an OpenSCENARIO 1.4 .xosc file.
• Parquet indexes for each split: 19,297 train, 2,320 validation and 2,866 test scenarios.
• Every scenario passed automatic quality gates before it was included.

Seven SimForge Worlds

A SimForge World is a release of a real place: an HD map (an OpenDRIVE road network surveyed to the kerb, and everything derived from it: lane graph, junction movements, a catalogue of places to stage a scenario) and a digital twin, the 3D environment Helios Engine renders. Seven ship in this release: San Ramon, phase 1; Di Rosa, San Francisco; Easterbrook Discovery School, San Jose; Belmont Research Center, Belmont; El Camino Road, Palo Alto; Yale Street, Palo Alto; Richmond Field Station, Richmond.

• Free with sign-in: simforge maps pull <map> --accept-license.
• Full 3D for Helios, plus OpenDRIVE and SUMO networks for any simulator that reads them.
• Vehicles, pedestrians and props come from the CARLA asset pack (CC BY 4.0); each download carries its notices.
• The maps are for non-commercial use under the SimForge account and download terms, and may not be redistributed.

Helios Engine

The runtime is the simulator and Helios Engine in one program: the same engine our cloud workers run. It is a closed binary, licensed under PolyForm Noncommercial 1.0.0, together with the SimForge account and download terms: you sign in to download it, and it reports usage.

• Linux x86_64 (Ubuntu 24.04 or newer for the CLI), macOS (Apple silicon; Intel for simulation and 2D) and Windows x64; the CLI also on Windows on Arm
• An NVIDIA GPU with 8 GiB or more, Vulkan 1.3, driver 570 or newer
• 2D renders, simulation, state-only RL, authoring and map downloads run on the CPU alone

How much GPU memory a map needs depends on its texture size. The CLI picks the size for your card:

• 8 GiB: 512 px textures on every public map, at most 4.5 GiB with six 1080p cameras
• 12 GiB: full textures on every map except San Ramon, phase 1
• 16 GiB: full textures everywhere (11.9 GiB peak)

Install the CLI, sign in, then install the runtime and a map:

curl -fsSL https://simforge.ai/install.sh | sh
simforge login --device
simforge runtime install --accept-license
simforge maps pull richmond-field-station --accept-license

The open toolkit

Everything that talks to the runtime and to SimForge's cloud is open source under Apache-2.0, at github.com/SimForgeinc/simforge-toolkit.

• simforge: the command line. Every command prints JSON, so people, scripts and AI agents use it the same way.
• SimForge Gym (the simforge-rl package): a Gymnasium environment over the runtime.
• simforge-cloud: a typed Python client for the cloud API.
• The scenario schema, sensor-rig presets, examples and agent skills.

pip install simforge-rl simforge-cloud

SimForge Gym

SimForge Gym (simforge-rl) is a Gymnasium environment: reset, step, observations, rewards. Episodes run in the runtime against traffic that reacts to the policy's car, at 50 Hz physics and 10 Hz decisions. A seed fixes the whole episode, so the same seed and the same actions give the same run. In sensor-only mode the policy receives rendered images and no engine state.

The baseline is PPO from the package's example, on the Richmond Field Station task set (24 tasks from 15 scenario templates), on one CPU host with no GPU:

The trained policy beats the scenarios' scripted driving on average and on 12 of 24 tasks, not on all of them. A second run with the same seed reproduced 11 of 11 shared checkpoints byte for byte. The method, the reward and the per-task table are on the SimForge Gym page.

SimForge Cloud: add --cloud when a laptop is not enough

SimForge Cloud is open in beta. The open stack and the cloud share one command line and one file format. Commands run on your machine by default; add --cloud and the same command runs on our GPUs, with a cost cap you set with --max-cost. Sign up and start with $10 in free credits; credit purchases aren't open during the beta. Each price is fixed before a run starts:

• 3D render: $1.00 per render.
• Open-loop evaluation: $0.25 per evaluation.
• Closed-loop evaluation: $1.00 per evaluation.
• Closed-loop evaluation of a model: $2.00 per evaluation.
• Transfer to another map: $0.04 per scenario.
• Recast (log to scenario): $1.00 per log.
• 3D asset generation: $1.00 per asset.
• 2D render: free.

With SimForge Cloud, Recast turns your own driving log into scenarios on our maps. RL training stays on your machine; the cloud does not train policies. Run in the cloud shows how to price, cap and download a run.

Built for AI agents

The CLI's JSON output, exit codes and errors are documented for programs as well as people. Coding agents can read the docs directly:

• simforge.ai/docs: install, guides and the full command reference.
• simforge.ai/llms.txt: an index of every docs page as Markdown.
• simforge.ai/docs/simforge-skills.zip: a skills pack that teaches an agent the CLI.

Licences

• Toolkit (CLI, Python clients, schema, examples): Apache-2.0.
• Runtime (simulator and Helios Engine): a closed binary under PolyForm Noncommercial 1.0.0, plus the SimForge account and download terms (sign-in to download, usage reporting).
• Dataset: CC BY-NC 4.0.
• SimForge Worlds (maps): non-commercial use under the SimForge account and download terms, no redistribution. Actor assets: CC BY 4.0.
• Commercial use of any part needs a commercial licence: contact@simforge.ai.