HY-LiYihan/Agentic-RSR-data

Agentic RSR Data

Companion experiment artifacts for Agentic RSR: Real-to-Sim-to-Real through Scene Reconstruction and Execution-Grounded Robot Policies.

• Paper (arXiv)
• Source code
• Project page
• Workflow and installation
• Artifact formats

Purpose And Scope

This repository distributes historical experiment artifacts and author-provided raw real-robot task videos associated with Agentic RSR. The project connects workspace-video reconstruction, Blender scene modeling, MuJoCo physics checks, and execution-guided robot policy development for Franka FR3 and AgileX Piper. It is an artifact collection rather than a standardized training dataset with predefined train/test splits. Use source revisions, timestamps and manifests to identify the provenance of historical artifacts.

Repository Layout

The existing archive was inspected at commit 4b1970c79b82407d2eceb54896b94720137e6509 on 2026-10-14. It contains 49,884 files before the addition of this card, organized under:

Path	Contents
summary/	Aggregate report, scene details and report-generation script
models/<model>/	Historical model-specific experiment bundles; inspected Claude Opus and GPT-6.1 bundles include a dsh/ directory
migration_checks/	Checks from the archive migration
migration_manifest.json	Migration status and job records
data/raw/frank/	9 raw FR3 real-robot task videos
data/raw/piper/	9 raw AgileX Piper real-robot task videos
data/raw/SHA256SUMS	SHA-256 checksums for the 18 original videos

Start with summary/REPORT.md and the matching experiment's report/configuration. Model directory names identify historical experiment configurations, not downloadable perception-model checkpoints. The aggregate archive spans multiple configurations; it should not be treated as a single paper result or as outputs from the current public checkout.

Artifact Formats

The following are formats used by the project, not a complete inventory of the files published in this dataset. Consult Files and versions for availability.

Artifact type	Associated files to preserve
Perception outputs	Sampled RGB frames, depth/confidence arrays, masks, camera parameters and inference manifests
Registered scenes	Metric depth/cameras, registration and delivery manifests, robot references and Blender scenes
Scene reconstruction	Blender files, rendered views, evaluations and source/input provenance
Physics and simulation	MuJoCo XML, referenced meshes/textures, export/derivation manifests and physics reports
Policy execution	Policies, run manifests, videos and reviews
Derived robot surfaces	Point clouds, metadata and upstream source/license records

Keep bundles together and preserve relative paths. Manifest paths may refer to the original checkout, and exported scenes can reference external files or symlinks. Relocation may require resolving those paths. A standalone .blend, XML or manifest is not necessarily a self-contained runnable scene.

Download And Use

Install the Hugging Face CLI. This dataset is public. Download the raw task videos with:

hf download HY-LiYihan/Agentic-RSR-data \
  --repo-type dataset --include 'data/raw/**/*.mp4' \
  --local-dir /path/to/agentic-rsr-data

The original videos are supplied at 1920x1080 with their original audio streams. Verify their integrity from data/raw/ with shasum -a 256 -c SHA256SUMS. Historical artifacts can be downloaded through Files and versions or --include filters. Pin an immutable commit when reproducing an experiment and follow the source repository's workflow for the matching artifact layout and code revision. Verify accompanying SHA256 records where provided before use. Model checkpoints are obtained separately from their official upstream sources; see the local inference guide.

Interpretation And Limitations

Historical experiment outputs do not establish that the current public local CUDA backend has been reproduced. Fresh local CUDA checkpoint inference remains unverified in the public code's validation record.

Reconstruction depth references are estimated depth, not independently measured physical ground truth. Registration or scene acceptance must be assessed using the relevant manifests and reports; file existence alone does not establish acceptance. Simulated physics results do not establish measured calibration or real-robot readiness. Hardware execution requires separate validation.

License And Provenance

Original project software and documentation are licensed under Apache-2.0. This does not assign Apache-2.0 to all files in this dataset. The task videos are author-provided recordings; robot assets, derived point clouds, model outputs and other third-party materials retain their respective rights and applicable terms. Public access does not grant additional copyright or trademark rights. Check the notices and provenance supplied with each artifact before reuse or redistribution. No dataset-wide license is asserted.

See the source repository's third-party records and asset distribution notes. Hugging Face hosting does not resolve missing upstream permissions or change the underlying licenses.

Citation

Please cite the paper:

@misc{li2026agenticrsr,
  title={Agentic RSR: Real-to-Sim-to-Real through Scene Reconstruction and Execution-Grounded Robot Policies},
  author={Li, Yihan and Feng, Yating and Sun, Shengjiu and Chen, Jianing and Ren, Hao and Yang, Bowen and Xu, Weisheng and Wu, Qiwei and Cheng, Hui and Xu, Renjing},
  year={2026},
  eprint={2610.10479},
  archivePrefix={arXiv},
  primaryClass={cs.RO},
  url={https://arxiv.org/abs/2610.10479}
}

For artifact questions or corrections, open an issue in the source repository.