# Bench2Dex v1.0.0 frozen ICL packages

This release contains the 26 existing `task_icl_v3` task ZIPs consumed by the
Bench2Dex RoboRSI integration. No ICL preparation is needed. All 33 source-release
files, including hidden files, README, audit, notices, catalog and ZIPs, are
preserved byte for byte. `RELEASE.md` and `MANIFEST.sha256` are new publication
metadata. The archived README describes the standalone evaluator; use this guide
for the shared dataset and RoboRSI.

## Contents and compatibility

- Tasks: 3, 6, 7, 8, 9, 12, 21, 22, 24, 26, 27, 32, 34, 42, 43, 44,
  51, 60, 61, 62, 64, 67, 73, 76, 79, 80.
- 26 ZIPs: 1,193,840,778 bytes (1.194 GB); 3,610 frozen payload files and
  26 task manifests; 1,234,937,299 expanded bytes.
- Each `packages/task_XX.zip` contains `manifest.json` and `payload/` with
  instructions, reference images, numerical trajectories and helper code.
- The ZIP catalog and hashes match `robots/bench2dex/icl_catalog.json` in
  [RoboRSI PR #80](https://github.com/SenseTime-FVG/RoboRSI/pull/80).
  The Agent/framework prerequisites and simulator requirements remain those
  documented in that integration. Download verification does not establish
  live behavior or success-rate equivalence.
- Baseline contract: Astra/max, Codex CLI 0.154.0, `baseline_4cam`,
  `calibrated_deployment_v2`. Package and container versions are independent.
  Native image: `ghcr.io/opensensenova/sensenova-roborsi-bench2dex@sha256:ba5253aa91e8e6888f0f0543626b9ace9c76dbe8559d7092caaa160e9b634a76`.

## Download and verify

Install `huggingface_hub` and run `hf auth login` with an account that has access
while this dataset is private. Do not place tokens in commands or source files.
Set `ROBORSI_PACKAGES_REVISION` to the **full HF commit SHA** from the publication
record in RoboRSI PR #80 (or this dataset's commit history). Record that SHA,
`bench2dex/v1.0.0`, code commit, and image digest with every evaluation.

Run this from the RoboRSI checkout using Python 3.9 or newer:

```python
import json
import os
from pathlib import Path
import subprocess
import sys
from huggingface_hub import snapshot_download

revision = os.environ["ROBORSI_PACKAGES_REVISION"]
assert len(revision) == 40 and all(c in "0123456789abcdef" for c in revision)
download_root = Path(snapshot_download(
    repo_id="sensenova/SenseNova-RoboRSI-Packages",
    repo_type="dataset",
    revision=revision,
    allow_patterns=["bench2dex/v1.0.0/*", "packages.json", "scripts/verify_package.py"],
    local_dir="work/roborsi-packages",
    token=True,
)).resolve()
package_root = download_root / "bench2dex/v1.0.0"
index = json.loads((download_root / "packages.json").read_text())
entry = next(p for p in index["packages"] if p["path"] == "bench2dex/v1.0.0")
subprocess.run([
    sys.executable, str(download_root / "scripts/verify_package.py"),
    str(package_root), "--manifest-sha256", entry["checksums_sha256"],
], check=True)
print(package_root)
```

The verifier checks the trusted release manifest, complete file set, all ZIP
hashes and sizes, task manifests and every payload hash. It reads ZIPs without
extracting them or executing their contents. Binary ZIPs are stored through
Git LFS; `snapshot_download` retrieves real content, not pointer files.
For a Git checkout, install Git LFS and run `git lfs pull` first.

## Use with RoboRSI

After installing the integration's documented Agent/framework dependencies,
stage a task Agent from the downloaded **version root**:

```bash
python -m benchmarks.bench2dex.agent_bundle --task 26 \
  --source work/roborsi-packages/bench2dex/v1.0.0 --output work/task26-agent
```

Use that same root for `benchmarks.bench2dex.run_suite --icl-source` and follow
the RoboRSI integration README for the remaining evaluation options. Choose a
fresh Agent output directory. The Agent verifies and consumes the original ZIP;
do not unpack, regenerate, rename, or edit it. Existing direct `--repo-id`
examples assume a standalone repository with `packages/` at its root, so use
`--source` for this shared repository layout. An absolute path resolved on your
own machine also works; no publisher-specific path is required.

For a single task, download `bench2dex/v1.0.0/packages/task_26.zip` at the same
revision and retain this relative layout. The Agent validates that task against
its checked-in catalog. The full release verifier requires all 26 ZIPs.

## Provenance, privacy and licensing

The original source release is dated 2026-10-12. Its public ZIPs already removed
session filenames from outer manifests while preserving model-visible payload
bytes. This upload makes no further changes. `AUDIT.json` records the earlier
integrity, heuristic privacy scan, sampled image review and offline selfchecks;
those checks are not a new simulator/model evaluation.

Source revisions: Bench2Dex `f96a8b2b4eb475483af66e9e03916b35bc43f1be`, IsaacLab
`37ddf626871758333d6ed89cf64ad702aef127d0`, `Bench2Dex/teleopdata`
`b195787c65046083e6a43776b67bdf1389dfa3eb`. Runtime assets are separate:
`Bench2Dex/Assets` at `bf65215f844d1fca30750e4cc70d7266650abf0a`.
Per-task `payload/SOURCES.md` files retain detailed attribution.

**Licensing remains pending as recorded in the original `LICENSE_NOTICE.md`.**
This private submission does not grant a new redistribution license or resolve
third-party permissions. Confirm owner-selected terms and applicable source
rights before public redistribution. Preserve the archived notice; if terms or
other package contents change, publish a new version directory.

中文：本版本完整保留原始 33 个发布文件及 26 个任务 ZIP，新增发布说明和校验清单。
下载时固定 HF 完整 commit SHA，并将 `bench2dex/v1.0.0` 目录传给 RoboRSI 的
`--source` / `--icl-source`。无需重新生成 ICL；不要将聚合仓库根目录直接传入。
私有仓库需要读取权限；公开分发前仍需确认原始许可声明中的待定事项。