pretty_name: SimCode4D DEV15
language:
- en
size_categories:
- n<1K
tags:
- robotics
- embodied-ai
- mujoco
- blender
- benchmark
- 3d
configs:
- config_name: dev15
data_files:
- split: validation
path: data/dev15.jsonl
SimCode4D DEV15
SimCode4D is the 15-scene development benchmark for Code2Worlds-2. It spans indoor, industrial, natural outdoor, service, and urban worlds (three scenes per domain). The table in Data Studio is a searchable scene index, not a leaderboard or a set of model outputs.
This repository holds benchmark data: prompts, 15 portable MuJoCo scene models with required mesh assets, 12 archived Blender authoring projects, and 18 approved scene previews. The benchmark evaluator, submission API, validators, and metric implementations live in the Code2Worlds-2 GitHub repository.
What is included
| Domain | Scenes | Per-scene preview files |
|---|---|---|
| Indoor | 3 | Not published |
| Industrial | 3 | Not published |
| Natural outdoor | 3 | NAT-01-rock-kick, NAT-02-hills-turn-stop, NAT-03-water-land |
| Service | 3 | SER-01-medicine, SER-02-airport-luggage, SER-03-hotel-drawer |
| Urban | 3 | Not published |
Only the natural outdoor and service scenes have published preview images. Each has three frames from an existing Blender animation (start, middle, end). We reviewed offscreen MuJoCo screenshots for the indoor, industrial, and urban scenes and omitted them because the simplified simulation geometry and temporary roof hiding did not represent the appearance of the authored worlds well. Their benchmark data remains available. A missing preview is represented as null in the index; it is not evidence that the case is absent. The available images are for visual inspection, not evidence of task success.
| Path | Contents |
|---|---|
data/dev15.jsonl |
15 prompts, domain and task labels, asset paths, and preview paths |
runtime/simcode4d_dev15_runtime.tar.gz |
Portable scene XML and required mesh assets for all 15 cases |
blender_sources/<case_id>/scene.blend |
12 available authoring projects |
previews/<case_id>/ |
Three approved Blender animation frames for six natural outdoor and service scenes |
THIRD_PARTY_NOTICES.md |
Asset provenance and third-party notices |
The original .blend sources for IND-01, IND-02, and IND-03 were not archived in the available benchmark handover. IND-01 and IND-02 point to the same missing pre-Module-1 source; IND-03 points to a separate missing source. Their compiled MuJoCo worlds are included. The index marks these three records as not_archived; a different Blender scene is not substituted for them.
Download
For the lightweight scene index, previews, and portable runtime:
hf download AIGeeksGroup/SimCode4D --repo-type dataset \
--include "data/*" --include "previews/*" --include "runtime/*" \
--local-dir SimCode4D
tar -xzf SimCode4D/runtime/simcode4d_dev15_runtime.tar.gz -C SimCode4D
The runtime_model field in each data/dev15.jsonl row points to a MuJoCo XML inside the extracted archive. For example:
import mujoco
model = mujoco.MjModel.from_xml_path(
"SimCode4D/benchmark/simcode4d_unified/artifacts/INS-01-agv-aisle/runtime_model.xml"
)
print(model.ngeom)
Download a Blender project separately when needed:
hf download AIGeeksGroup/SimCode4D blender_sources/INS-01-agv-aisle/scene.blend \
--repo-type dataset --local-dir SimCode4D
Opening authoring projects may require Blender and the original scene-generation dependencies. The MuJoCo runtime archive is the portable path for inspecting and executing the benchmark worlds.
Evaluation
Use the GitHub benchmark guide to run the reference integrity check or score a generated world. It defines WorldBuildPass@1, TaskOverlayPass@1, entity binding F1, task success, mechanics, kinematics, contact/friction, safety, and robust task success. The GitHub release retains the frozen verification metadata needed for reproducible scoring; this data repository presents the scenes and their previews without repeating those metadata tables or shipping a second evaluator.
The split is development/validation, not an unseen test set. The reference matrix checks benchmark self-consistency; it is not a Code2Worlds-2 method score.
Provenance and use
The portable runtime data comes from the public Code2Worlds-2 development suite. The Blender projects are the files named in that suite's source records; all 12 available files were checked against the source records before upload. Third-party model notices are in THIRD_PARTY_NOTICES.md. No repository-wide license is asserted here; review the notices and the applicable project terms before reuse.