Real-world driving scenarios you can run closed loop, with reactive traffic and camera and lidar rendering on 3D digital twins of real places.
To our knowledge, the largest public set of real-world driving scenarios that run closed loop with reactive agents and camera and lidar rendering on 3D digital twins of real places (as of October 2026). See the comparison.
What it is
SIMLOG-25k holds 24,483 driving scenarios on 7 SimForge Worlds.
Every scenario runs closed loop in SimForge, so a policy can take the wheel while the traffic around it reacts: for closed-loop evaluation, and for reinforcement learning in SimForge Gym.
Each one was simulated again on the released runtime and kept only if it runs, its motion is physically plausible, it has no unlabelled collision and its key vehicles obey signals and stop lines.
How it compares
| Dataset | Real-world scenarios | Closed loop, reactive agents | Sensor rendering in closed loop | 3D digital twin of a real place | Count, in each dataset's own unit | Licence |
|---|---|---|---|---|---|---|
| SIMLOG-25k | Yes | YesReactive traffic, SimForge Gym | YesCamera and lidar, Helios Engine | Yes7 real places | 24,483 scenarios | CC BY-NC 4.0, gated |
| World Engine | Yes | Yes | Yes3DGS | Yes3DGS of real places | 288 long-tail scenarios1 | Not stated |
| HUGSIM | YesWith inserted actors | PartlyReactive (IDM) on nuScenes only | PartlyCamera only, 3DGS | YesReconstructions | 70+ sequences, 400+ scenarios2 | Code MIT; data licence not stated |
| NAVSIM v1 / v2 | Yes | PartlyNon-reactive pseudo-simulation | Partlyv2: pre-rendered 3DGS only | NonuPlan maps | About 12,000 navtest samples; navhard 450 / 5,4623 | nuPlan terms |
| nuPlan | Yes | YesReactive agents (IDM) | No | NoHD map | 1,282 h of driving, in 15 s scenarios4 | CC BY-NC-SA 4.0 + nuPlan terms |
| interPlan | YesHand-modified | YesReactive agents (IDM) | No | NonuPlan maps | 80 scenarios (335 in all)5 | nuPlan terms |
| Waymo Open Motion + Waymax | Yes | YesIDM or log replay | NoBoxes only | NoHD map vectors | 103,354 segments of 20 s6 | Waymo non-commercial |
| Bench2Drive | NoSynthetic | Yes | YesGame engine | NoSynthetic towns | 220 evaluation routes; 13,638 training clips7 | CC BY-NC-ND |
| CARLA Leaderboard 2.0 | NoSynthetic | Yes | YesGame engine | NoSynthetic towns | 90 training and 20 validation routes8 | MIT / CC BY |
| nuScenes | Yes | No | NoLogged sensors only | NoHD map | 1,000 scenes of 20 s9 | CC BY-NC-SA 4.0 |
| Argoverse 2 Motion | Yes | No | No | NoHD map | 250,000 scenarios of 11 s10 | CC BY-NC-SA 4.0 |
Public driving-scenario datasets compared on what closed-loop sensor simulation needs: real-world scenarios, reactive closed loop, sensor rendering and a 3D model of a real place. Counts use each dataset's own unit. Sources checked October 2026.
- World Engine: arXiv 2606.19836
- HUGSIM: arXiv 2412.01718
- NAVSIM v1 / v2: arXiv 2406.15349; arXiv 2506.04218
- nuPlan: arXiv 2403.04133
- interPlan: github.com/mh0797/interPlan; arXiv 2404.07569
- Waymo Open Motion + Waymax: waymo.com/open/about; github.com/waymo-research/waymax
- Bench2Drive: arXiv 2406.03877
- CARLA Leaderboard 2.0: arXiv 2412.09602
- nuScenes: arXiv 1903.11027
- Argoverse 2 Motion: argoverse.org/av2.html
What's in it
Each scenario is a folder:
- The scenario. A SimForge scenario document: the map, the road users, their routes and what triggers what.
- The trace. The simulation result: every road user's pose every 20 ms, as the released runtime ran it.
- OpenSCENARIO. The same scenario as an ASAM OpenSCENARIO 1.4 file, for other simulators.
Positions only: no video and no sensor data. You render a scenario yourself, in Helios Engine.
- Splits. train, validation and test: 19,297, 2,320 and 2,866 scenarios.
Get it
CC BY-NC 4.0: free for research, evaluation and other non-commercial use, with attribution. The dataset is gated: sign in to Hugging Face and accept the terms on its page first. For a commercial licence, write to contact@simforge.ai.
Load the index with the datasets library:
from datasets import load_dataset
ds = load_dataset("SIMFORGE-AI/simlog-25k") # train, validation, testEach row names its scenario's files. To work with the files, download the whole dataset:
hf download SIMFORGE-AI/simlog-25k --repo-type dataset --local-dir simlog-25kRun it in SimForge
With the CLI and runtime installed (Quickstart), pull the World a scenario names, then simulate it and render it in Helios:
simforge maps pull el-camino-road --accept-license
simforge simulate <scenario>/document.json --out sim
simforge render <scenario>/document.json --mode 3d --out framesOr train a policy on a split in SimForge Gym (RL reference):
python -m simforge_rl.examples.ppo --export simlog-25k \
--split train --runtime --max-open 16 --out run/