YuMoool/astra-robodojo-rollouts

Astra RoboDojo Evaluation Records

Rollout records from the evaluations in GPT 6 Astra as an Embodied Policy, by Jiayi Su, Yixin Zheng, Mi Yan, Li Yi, Zhizheng Zhang, and He Wang. This archive includes action proposals, executed actions, observations, robot states, model-provided explanations and reasoning summaries, and metadata for reproducing the evaluation settings, together with a reader and documentation.

• Report
• Public controller source
• Data schema and alignment
• Reproduction guide and limitations

The release contains the 100 selected final evaluation records: ten tasks, five fixed cases per task, two methods. hybrid is π₀.₅ + GPT 6 Astra; direct is GPT 6 Astra（direct）, with no π₀.₅ inference. All records use gpt-6-astra and xhigh. These are recorded evaluation runs, not a curated training split. It does not include discarded retry attempts, earlier pilot runs, RoboLab, or a separately executed π₀.₅-only baseline.

Method	Cases	Native-complete trajectories	Successes	Success rate	Mean native Score × 100
π₀.₅ + GPT 6 Astra	50	50	24	48%	62.60 (50 scores)
GPT 6 Astra（direct）	50	48	13	26%	37.8125 (48 scores)

The two Direct simulator-RPC idle timeouts were explicitly adjudicated as evaluation failures. They retain their executed prefix, native_complete=false, native_success=null and native_score=null; they are not native-terminal trajectories. Both remain in the success-rate denominator. Missing scores are not filled with zero. The subset was chosen using official π₀.₅ success rates; these are not full-benchmark leaderboard results.

Quick start: read without running a model or simulator

Python 3.10+ is required. NumPy and Pillow are enough to read downloaded files; huggingface_hub is only needed for downloading. No PyTorch, JAX, GPU, simulator, Codex login or paid inference is needed for reading, visualization or recounting.

pip install -r requirements.txt

Download just the metadata, reader, and folded-clothes core records first:

from huggingface_hub import snapshot_download

root = snapshot_download(
    repo_id="YuMoool/astra-robodojo-rollouts",
    repo_type="dataset",
    local_dir="./astra-robodojo-rollouts",
    # For reproducibility, set revision to the full commit SHA you downloaded.
    allow_patterns=["*.json", "*.jsonl", "*.md", "*.py", "requirements.txt", "LICENSE",
                    "episodes/*/fold_clothes/*.core.zip"],
)

Keep local_dir set: it produces a self-contained directory, while the reader intentionally rejects file links that escape its download directory.

Run python example.py /path/to/download --method hybrid --task fold_clothes, or import the reader (put the downloaded directory on Python's import path):

from reader import Dataset

dataset = Dataset("/path/to/download")
record = dataset.select(method="hybrid", task="fold_clothes")[0]
with dataset.open(record["episode_id"], verify=True) as episode:
    print(episode.metadata["case"], episode.metadata["outcome"])
    for decision in episode.decisions:
        obs = decision.observation(images="preview")
        proposal = decision.pi05_proposal  # raw_actions/actions, each (50,14); None for direct
        target_trace = decision.eef_targets  # per-step EEF targets and IK diagnostics
        actual_actions = decision.executed_actions
        reason = decision.requested_response.get("reason")

images="preview" loads the actual saved decision-time previews, not a new resizing of source frames. To read full-resolution RGB at any control step, also download the corresponding .observations.zip and use episode.raw_observation(t). Avoid requesting all raw images unless needed. ZIP members are read directly, without extraction or Python pickle execution.

Files and download tiers

Each episode is independently downloadable:

manifest.json                     selected episode index, results, seeds, sizes and SHA256
episodes.jsonl                    standard JSONL catalog for the Hub dataset viewer
validation.json                   independent validation and numerical recount receipt
pairing.json                      original50-pair initial-state/configuration audit, paths removed
reader.py / example.py            lightweight Python reader and usage example
prepare_case.py                   write selected-case env/config and optionally verify upstream files
SCHEMA.md / REPRODUCE.md           field definitions, alignment and reproduction requirements
episodes/{hybrid|direct}/{task}/
  {episode_id}.core.zip            decisions, action/state arrays, preview images, prompts, provenance
  {episode_id}.observations.zip    original lossless NPZ for every control-step observation
  {episode_id}.receipt.json        export source/output identities

Core bundles contain all low-dimensional states and executed actions, candidate π₀.₅ chunks, structured GPT decisions, requests, correction traces, service calls, public model messages/reasoning summaries and decision-time image previews. Optional observation bundles retain all three original 640×480 RGB arrays, 14D proprioception, EEF poses, instruction and remaining step count per control step. All raw NPZ bytes are identical to the source archive, verified by SHA256. There is no video transcoding, image quantization or action rounding.

Data sizes and hashes are authoritative in manifest.json; the run directory sizes included redundant copies and are not the release size. Core bundles omit duplicate camera arrays inside proposal files, retaining the original numerical proposal arrays and their source-file hashes. Observation/state and decision identifiers provide the join back to those inputs.

Validation and reported numbers

Download all core ZIPs, then run:

python validate.py /path/to/download --require100 --output validation-local.json

This checks checksums, schema, episode/decision alignment, action slices, candidate-prefix execution, native/adjudicated outcomes, paired seeds/layouts, and recalculates the reported success and Score values. The security check rejects known credential patterns, private host paths and internal gateway addresses. With all raw observation bundles downloaded, add --full to verify every raw frame hash and its state-array agreement. This reads all archived observations and takes substantially longer. The release receipt states which checks ran.

Token totals are the controller-recorded usage, including reused cached input. Cached input is a subset of input, and reasoning output is a subset of output; do not add either subset again. These are usage records, not an invoice or a reconstruction of provider billing.

What the public model text does and does not contain

reason and assessment are the controller's explicit structured explanations. agent_visible_events.jsonl contains visible assistant messages and, when returned by the interface, public reasoning summaries. These are not private chain-of-thought, nor a claim to the complete internal model context. Model compaction, hidden state, native-tool stdout and account/runtime events are not redistributed. Service request/result records are preserved with paths normalized. Do not treat every simulator frame as an image actually sent to the model.

Rights, privacy and intended use

Authored rollout records and documentation are released under CC BY 4.0; the reader/validation/example code is MIT. Third-party simulator content, models and dependencies retain their own rights. No checkpoints, USD/mesh/texture assets, simulator installation, credentials, account databases, private annotations, or raw unfiltered model-session logs are included. Scene configuration references and code provenance are supplied for alignment, not as a redistribution of the simulator itself. See LICENSE-DATA.md and LICENSE.

These data support behavioral analysis, verification of recorded results, and reproduction of evaluation settings. They do not establish hardware safety, real-time deployment or deterministic reproduction of a hosted model's actions. The two methods share case definitions, not identical decision counts, token budgets or chunk lengths. Five cases per task limit statistical precision.

Citation

@misc{su2026astra,
  title = {GPT 6 Astra as an Embodied Policy},
  author = {Jiayi Su and Yixin Zheng and Mi Yan and Li Yi and Zhizheng Zhang and He Wang},
  year = {2026},
  url = {https://anonymous-report-421.github.io/public-website/},
  note = {RoboDojo evaluation records: https://huggingface.co/datasets/YuMoool/astra-robodojo-rollouts}
}

Please also cite RoboDojo and π₀.₅/OpenPI when using their benchmark or policy. Pin the archive revision in derived results and report any excluded records.