---
pretty_name: SimCode4D DEV15
language:
  - en
tags:
  - embodied-ai
  - robotics
  - mujoco
  - benchmark
configs:
  - config_name: dev15
    data_files:
      - split: validation
        path: data/dev15.jsonl
---

# SimCode4D: Code2Worlds-2 development benchmark

SimCode4D DEV15 is the public, portable 15-case development benchmark for the
Code2Worlds-2 paper. It covers indoor, industrial, natural outdoor, service,
and urban environments, with three cases in each domain. The benchmark tests
whether a generated world builds and whether its physical behavior satisfies
the registered task, integrity, mechanics, kinematics, contact, friction, and
safety checks.

The `validation` split in the viewer is a **case index**. Each row points to
the complete frozen specification, gold predicates, physical profile, task
confirmation, and runtime artifacts in this repository. It is not a table of
model outputs or experiment results.

## Contents

| Path | Contents |
| --- | --- |
| `data/dev15.jsonl` | Searchable index of all 15 development cases |
| `benchmark/simcode4d_unified/config/` | Suite registry, seed protocol, and frozen hashes |
| `benchmark/simcode4d_unified/cases/` | Case specifications, gold predicates, physics profiles, and task confirmations |
| `benchmark/simcode4d_unified/artifacts/` | Frozen M1 and executable MuJoCo runtime models and overlays |
| `benchmark/simcode4d_unified/assets/` and `vendor/` | Runtime meshes and third-party model assets |
| `benchmark/simcode4d_unified/probes/`, `validators/`, and `tools/` | Reference probes, independent validators, and scoring tools |
| `requirements-benchmark.txt` | Python dependencies for the public benchmark runtime |

The runtime package is copied from the public
[Code2Worlds-2 repository](https://github.com/AIGeeksGroup/Code2Worlds-2/tree/ca35f25cda19aadf1d211ae3e55e390905507f1d/benchmark/simcode4d_unified)
at commit `ca35f25cda19aadf1d211ae3e55e390905507f1d`. Paths are preserved so
that the original registry, hashes, and evaluator work without rewriting the
cases. The case index is an additional convenience file.

## Download and verify

Use Python 3.10. From the downloaded dataset root:

```bash
hf download AIGeeksGroup/SimCode4D --repo-type dataset --local-dir SimCode4D
cd SimCode4D
python -m pip install -r requirements-benchmark.txt
python benchmark/simcode4d_unified/check_public_release.py
```

The integrity check verifies the frozen references and loads all 15 compiled
M1 worlds and runtime models with MuJoCo. The expected output has `"passed":
true` and `"case_count": 15`.

To test the submission interface without generating a new world:

```bash
python benchmark/simcode4d_unified/tools/make_reference_smoke.py \
  --output /tmp/simcode4d-reference-package
python benchmark/simcode4d_unified/tools/score_submission.py \
  --submission-root /tmp/simcode4d-reference-package \
  --output /tmp/simcode4d-reference-score \
  --case-ids SER-03-hotel-drawer --physics-seeds 401
```

This is a reference-interface smoke test, not a generated-method score. See
the [benchmark README](benchmark/simcode4d_unified/README.md) for the full
matrix, method submission format, and metric definitions.

## Scope and provenance

The registered reference matrix contains 15 cases × 3 deterministic rebuild
slots × 5 physics seeds = 225 executions. The three rebuild slots do not
represent independent generated worlds; there are 75 distinct case-physics
trials. The reference-probe results described in the benchmark README check
benchmark self-consistency and must not be presented as a Code2Worlds-2 model
score or as a held-out leaderboard.

This portable release excludes archived Blender authoring scenes, ablation
configurations, completed run logs, leaderboards, and demo videos. The
original authoring sources for three indoor scenes were unavailable. Three
cases have path-only asset portability repairs; see
[PORTABILITY.md](benchmark/simcode4d_unified/PORTABILITY.md). The source-name
binding metric requires authoring-source catalogs outside this compact
release; the included runtime binding check uses compiled model interfaces.

No repository-wide license was specified in the source release. Third-party
model assets retain their license and notice files under `vendor/`; review
those notices before reuse beyond the benchmark runtime.