Dream.exe: Can Video Generation Models Dream Executable Robot Manipulation?

Rui Zhao^1,*, Kaiming Yang^1,*, Jifeng Zhu^1,†, Siyang Chen^1,†, Ziqi Wang¹, Weijia Wu¹, Kevin Qinghong Lin², Heng Wang³, Mike Zheng Shou^1,‡

¹Show Lab, National University of Singapore   ²University of Oxford   ³Tencent

_{^*Equal contribution   ^†Equal contribution (second authors)   ^‡Corresponding author}

Paper · Code

Dataset overview

Dream.exe asks whether a generated manipulation video can be converted into a 3D robot trajectory, executed in simulation, and judged by whether the imagined task actually succeeds. This repository contains versioned benchmark inputs and released run data used to study that question; the implementation lives in the separate Dream.exe code repository.

The benchmark currently comprises 101 curated robot-manipulation task instances spanning single-object manipulation, multi-object interaction, and long-horizon multi-stage tasks. These instances were constructed by the Dream.exe authors on top of RoboCasa simulation environments. Their task definitions, scene configurations, initial states, camera settings, generation inputs, protocols, and references were selected, modified, or produced specifically for this benchmark; they are not an unchanged copy of a RoboCasa dataset. Released results pair model videos with reproducible run specifications and the resulting trajectories, actions, execution traces, videos, diagnostics, and evaluation records.

Dataset contents

Part	Contents
bench/	Curated benchmark cases, protocols, provenance, and the required simulator-asset closure
results/	Published run specifications, model-video bundles, and their referenced evaluation artifacts

Repository structure

.
├── bench/
│   ├── bench.json
│   ├── collections.json
│   ├── protocol/                    # stage and evaluation contracts
│   ├── sources/                     # pinned runtime provenance and simulator assets
│   └── cases/<uid>/
│       ├── env/                     # versioned scene, state, camera, and initialization data
│       ├── generation/              # first frame and generation prompts
│       ├── protocol.json            # optional case-specific protocol values
│       └── references/              # ground-truth video, action, and depth
└── results/
    ├── runs/<model>/<variant>/run.json
    ├── videos/<uid>/<model>/<variant>/
    │   ├── video.mp4                # original model output
    │   ├── preprocessed.mp4         # pipeline input when present
    │   └── video.json               # provenance and content identities
    └── experiments/<uid>/<model>/<variant>/
        ├── request.json
        ├── resolved_config.json
        ├── result.json              # artifact manifest and completion record
        ├── trajectory/              # extracted 3D trajectories
        ├── action/                  # executable action sequence and metadata
        ├── execution/               # rollout video, summary, and traces
        ├── evaluation/              # task-success and executability records
        ├── intermediate/            # stage diagnostics used by the release
        └── logs/

bench/: curated evaluation inputs

bench/ defines the evaluation problem for a published revision. A case ID identifies one benchmark task instance. Within a pinned revision, the scene, simulator state, camera, generation input, protocol, and references remain fixed. Global protocol documents define generation, video-to-trajectory, action, execution, and evaluation behavior; a case-level protocol.json records only exceptional values. Ground-truth video, action, and depth are references and must not be treated as generated-model results.

The benchmark also contains the RoboCasa/RoboSuite simulator-asset closure needed to reconstruct its versioned scenes. It does not contain the upstream RoboCasa source episode dataset.

results/: released runs and artifacts

results/runs/ is the published run matrix consumed by Dream.exe reproduction commands. It includes standard and enhanced variants for the evaluated video-generation/world-action models, plus reference runs with and without ground-truth depth.

results/videos/ stores model outputs separately from execution videos. A video bundle can be staged or published before a corresponding experiment artifact is available, so directory presence alone must not be interpreted as experiment completion. results/experiments/ contains released evaluation artifacts and manifests. Historical partial runs remain partial; missing artifacts are not synthesized. Treat a pinned dataset revision as immutable publication data and write locally recomputed outputs to a separate workspace.

Download and reproduce

Install the Hugging Face CLI and download the repository into the code repository's data/ directory:

python -m pip install -U huggingface_hub

hf download kaimingyang/Dream.exe \
  --repo-type dataset \
  --local-dir data

This produces data/bench/ and data/results/, matching the default Dream.exe workspace. Follow the full benchmark guide for environment validation and reproduction commands.

For a reproducible experiment, pin the exact dataset commit:

hf download kaimingyang/Dream.exe \
  --repo-type dataset \
  --revision <DATASET_COMMIT> \
  --local-dir data

Versioning and updates

This dataset is versioned through Hugging Face commits. New cases or results may be added, corrected entries may be replaced, and withdrawn entries may be removed from the latest revision. Always record the dataset commit SHA in papers, reports, and derived artifacts. Removing a path from the latest revision does not erase it from older revisions.

Limitations and responsible use

• The benchmark evaluates simulated robot manipulation; it does not establish real-robot safety or deployment readiness.
• Generated videos and historical results can be incomplete or unsuccessful by design. Use result.json, manifests, and evaluation records rather than inferring completion from directory presence.
• Some video provenance records currently report rights.status: unknown. Public availability does not itself grant permission to reuse an upstream model output for every purpose; review the per-artifact metadata and the applicable upstream terms.
• Avoid publishing conclusions without pinning a dataset revision and preserving the benchmark/result distinction above.

Licenses and attribution

There is no single blanket license for every file in this dataset. Dream.exe-authored code is licensed separately under Apache-2.0 in the code repository. RoboCasa assets in the bundled simulator closure are provided under CC BY 4.0, and the included RoboSuite/RoboCasa license and attribution files must be retained. Generated videos, model outputs, and other third-party-derived artifacts may be governed by separate upstream terms. This repository does not grant additional rights beyond those terms.

See bench/sources/simulator-assets/licenses/ and the code repository's third-party inventory before redistribution or commercial use.

Citation

@article{zhao2026dreamexe,
  title   = {Dream.exe: Can Video Generation Models Dream Executable Robot Manipulation?},
  author  = {Zhao, Rui and Yang, Kaiming and Zhu, Jifeng and Chen, Siyang and Wang, Ziqi and Wu, Weijia and Lin, Kevin Qinghong and Wang, Heng and Shou, Mike Zheng},
  journal = {arXiv preprint arXiv:2606.04811},
  year    = {2026}
}