RoboCasa-Shift Formal Evaluation Records

Project links: Paper (arXiv) · Code (GitHub)

This dataset contains the episode-level records for the formal RoboCasa-Shift evaluation reported in the paper RoboCasa-Shift: Stage-Aware Robustness Evaluation of Vision-Language-Action Policies in Long-Horizon Household Manipulation.

The evaluation covers four clean-qualified RoboCasa365 Composite-Seen tasks with the GR00T N1.5 Composite-Seen policy. For each of 20 matched base episodes per task, it records one clean rollout and 12 single-factor shifted rollouts across illumination, camera orientation, robot base initialization, and Gaussian blur at three severity levels. The release contains 1,040 validated episode records (260 per task).

The records include official task outcomes, clean/shift condition identity, pair identifiers, seeds, simulator-state fingerprints, perturbation parameters, and StageTracker milestone/progress fields. StageTracker is diagnostic; it does not replace the official task-success criterion.

Load the records

from datasets import load_dataset

dataset = load_dataset("HaochengWang23/robocasa-shift", "formal_evaluation")
print(dataset)
print(dataset["PreSoakPan"][0])

Each task is a split in the formal_evaluation configuration. The JSONL files can also be downloaded directly from the formal_evaluation/tasks/ directory.

Files

Path	Contents
formal_evaluation/tasks/<task>/episodes.jsonl	260 sanitized formal episode records for each task
formal_evaluation/tasks/<task>/validation_report.json	Pairing and record validation summary for each task
formal_evaluation/task_freeze.json	Frozen task-suite metadata
formal_evaluation/DATA_MANIFEST.csv	Record counts and SHA-256 checksums for the four JSONL files

The release excludes model weights, RoboCasa source datasets, simulator state files, raw rollout videos, and machine-specific artifact paths. The included episode records are sufficient for the published tabular re-analysis; they are not sufficient to rerun the policy without the separately licensed model and simulator assets.

Provenance and use

These are the sanitized formal evaluation records distributed with the RoboCasa-Shift code repository. The paper reports the same 1,040-rollout formal evaluation. See the repository's benchmark protocol and data and model notes for the evaluation protocol, data boundaries, and upstream asset requirements.

This is one evaluation of one policy on a fixed task suite. Results should not be interpreted as a general ranking of VLA models or as evidence that a perturbation improves performance in general. A failure-to-success pair is an episode-level transition under the matched protocol.

Citation

@article{wang2026robocasa_shift,
  title = {RoboCasa-Shift: Stage-Aware Robustness Evaluation of Vision-Language-Action Policies in Long-Horizon Household Manipulation},
  author = {Wang, Haocheng and Liu, Youquan and Guo, Weiyu and Zhang, Aokai and Wu, Donghang and Liu, Hongyu and Zhang, Sha and Yang, Weidong and Fei, Ben},
  year = {2026},
  eprint = {2610.27861},
  archivePrefix = {arXiv},
  primaryClass = {cs.RO},
  url = {https://arxiv.org/abs/2610.27861}
}

License

The formal evaluation records in this dataset are licensed under the Creative Commons Attribution 4.0 International License (CC BY 4.0). Please cite the paper when using these records. This dataset license applies to the evaluation records in this repository; the RoboCasa-Shift code repository is separately licensed under MIT, and third-party software, models, and upstream datasets remain under their own terms.