sensenova/SenseNova-RoboRSI-Packages

SenseNova RoboRSI Packages

Versioned experience and in-context learning (ICL) packages for benchmarks integrated into SenseNova RoboRSI. This dataset repository stores the resources consumed by the evaluation framework. Simulator environments and benchmark Docker images are distributed separately.

Available packages

Benchmark	Package version	Condition	Tasks	Package root
DexJoCo (dexjoco)	v1.0.0	rand_full	11	dexjoco/v1.0.0/
SIMPLE (simple)	v1.0.0	simple_native_training_icl_v1	5	simple/v1.0.0/
Bench2Dex (bench2dex)	v1.0.0	task_icl_v3	26	bench2dex/v1.0.0/
RoboDojo (robodojo)	v1.0.0	public_training_frozen	42 base / 54 mapped	robodojo/v1.0.0/
LIBERO-Long++ (libero-long-plus-plus)	v1.0.0	efficiency_aligned_v12	12 mapped / 12 packages	libero-long-plus-plus/v1.0.0/
LIBERO-Ultra-Long (libero-ultra-long)	v1.0.0	efficiency_aligned_v12	9 mapped / 9 packages	libero-ultra-long/v1.0.0/
LIBERO-PRO (libero-pro)	v1.0.0	pro_persistent_memory_v18	80 mapped / 40 packages	libero-pro/v1.0.0/
RoboCasa365 Target50 (robocasa365)	v1.0.0	target50_same_task_seed0	43 packages / 305 of 340 cells	robocasa365/v1.0.0/
BEHAVIOR (behavior)	v1.0.0	official_21d_icl	100	behavior/v1.0.0/

packages.json is the machine-readable release index. Paths in that index are relative to this repository root. Published version directories are immutable; changes to an experience package require a new version directory. Pin the full Hugging Face repository commit SHA when running evaluations.

Repository layout

README.md
LICENSE
packages.json
GUIDE.zh-CN.md
scripts/verify_package.py
dexjoco/v1.0.0/
  README.md
  LICENSE
  LICENSE-MIT
  THIRD_PARTY_NOTICES.md
  third_party/
  MANIFEST.sha256
  knowledge_packages.json
  rand_full/<task>/manifest.json
  rand_full/<task>/payload/...

The DexJoCo release preserves all 292 source files byte for byte (34,781,416 bytes). Its task directories contain 279 files (34,693,512 bytes), including 154 JPEG images and 27 NumPy files. All 181 image/NumPy files use Git LFS. The original MANIFEST.sha256 covers 291 package files, excluding itself. The catalog pins 11 task manifests and 268 payload files.

Download a pinned release

Install huggingface_hub. Authenticate with hf auth login if the repository is private and your account has access. Replace the placeholder below with a full commit SHA from this repository's history.

from pathlib import Path
from huggingface_hub import snapshot_download

download_root = Path(snapshot_download(
    repo_id="sensenova/SenseNova-RoboRSI-Packages",
    repo_type="dataset",
    revision="<FULL_COMMIT_SHA>",
    allow_patterns=["dexjoco/v1.0.0/*", "scripts/verify_package.py", "packages.json"],
    local_dir="./roborsi-packages",
))
package_root = download_root / "dexjoco" / "v1.0.0"
assert (package_root / "rand_full").is_dir()

huggingface_hub downloads the actual LFS file contents. For a Git checkout, install Git LFS and run git lfs pull to populate binary files.

Pass the package root, not the repository/download root, to the companion DexJoCo evaluator. From that evaluator's code checkout, for example:

python scripts/fetch_knowledge.py --local-dir /absolute/path/to/roborsi-packages/dexjoco/v1.0.0
python scripts/verify_release.py --require-knowledge

Alternatively, pass the same package root to run_frozen.py --knowledge-root. The archived package README describes the original standalone release layout; its direct fetch_knowledge.py --repo-id ... example must not be applied to this aggregate repository unless that downloader supports a package subdirectory. Use --local-dir as shown above. The catalog's knowledge/rand_full/... paths describe installation into the companion evaluator; the package-local path is rand_full/.... Keep frozen payload files and task manifests unchanged.

To verify the downloaded package, run the following from the same working directory as the download example. The trusted manifest hash below is the checksums_sha256 recorded for DexJoCo v1.0.0 in the release index:

python roborsi-packages/scripts/verify_package.py roborsi-packages/dexjoco/v1.0.0 --manifest-sha256 024dca42f2424cb77f9fdba433f6d6447fe9b8c6efac082febae0c7f3c9b7935

This checks hashes; it does not execute any bundled Python policy/helper code. Running the benchmark additionally requires the compatible evaluator and simulator environment. Download/hash verification is not an evaluation smoke test.

Provenance and licenses

DexJoCo simulator commit: 8d23b0fab23b17a58c4b55f3942e17013aaf8267. Source demonstration dataset: DexJoCo/DexJoCo-Datasets-Raw, revision 125df5c1019e97503929ef1a0ad8f90373436afa. See the package README, manifests, and per-task payload/SOURCES.md for preparation model, protocol, evidence, and source details. These packages apply to rand_full; do not relabel them as rand_obj packages.

Licenses are component-specific. Retain the package's LICENSE, LICENSE-MIT, THIRD_PARTY_NOTICES.md, third_party/, and per-task source records. In particular, the bundled notices identify GeoRT-derived material with CC BY-NC noncommercial terms. This aggregate repository does not relicense its contents.

Contribution guide

See the Chinese organization and upload guide, and the team's RoboRSI ICL material upload guide. Binary assets must be tracked with Git LFS. Benchmark IDs follow the benchmark Docker image naming convention, including dexjoco for DexJoCo. Each benchmark versions its packages independently; Docker image versions and package versions are separate identifiers and their compatibility must be documented.

SIMPLE v1.0.0

See the SIMPLE release guide for download, verification, task mapping, and integration details. Five frozen task packages are included: XMovePick, BendPick, Handover, Tabletop, and PickBetweenTables. XMoveBendPick has no ICL package in this version. These contain 2,385 frozen payload files plus their five original manifests (about 200 MB), including training references, images/videos, numerical trajectories, robot metadata, and helper scripts. No learned policy weights or evaluation episode logs are included.

Download simple/v1.0.0/*, packages.json, and scripts/verify_package.py at a fixed full repository commit. Use the SIMPLE entry's checksums_sha256 with the shared verifier. Package-local task roots are tasks/<Task>/; do not use DexJoCo-specific evaluator commands for SIMPLE. The archived package README is retained byte for byte; RELEASE.md contains the current publication guide. See SIMPLE component notices for provenance and data-license limitations. Hash verification does not run a simulator or model.

Bench2Dex v1.0.0

See the Bench2Dex release guide for pinned download, verification and RoboRSI Agent staging. The 26 frozen task ZIPs total 1.194 GB and contain 3,610 payload files. All 33 original release files, including hidden files and notices, are preserved byte for byte. The shared verifier supports this ZIP catalog without extracting or executing payloads; existing DexJoCo and SIMPLE directory catalogs remain supported.

Download bench2dex/v1.0.0/*, packages.json and scripts/verify_package.py at a full repository commit, then verify using the Bench2Dex index entry's checksums_sha256. Pass the version root to RoboRSI agent_bundle --source or run_suite --icl-source; this aggregate repository is not compatible with the standalone root-level --repo-id download example. Official benchmark assets and the simulator image are separate. The original pending-license notice remains applicable; this private submission does not establish public redistribution rights.

RoboDojo v1.0.0

See the RoboDojo usage guide. The release contains 27 complete sealed tar.gz experience packages: 21,387 original payload files and 2,521,855,428 uncompressed payload bytes, with 2,418,619,576 archive bytes. The catalog maps 42 original tasks and explicitly transfers base knowledge to 12 generalization variants; it includes no variant-specific demonstrations. These are historical public-training references, not weights or evaluation runs.

Download robodojo/v1.0.0/*, packages.json and scripts/verify_package.py at a full repository commit. Verify the version root using trusted manifest SHA256 7f65416beb515b888f253be2da8ae4a27fa868359d09200659000f4a97ef00d5. The shared verifier checks tar archive, package-manifest and payload hashes without extracting or executing the frozen helper code. The sensenova_roborsi XPolicyLab policy installs these exact archives and uses the standard Model/reset/update_obs/get_action and eval.sh entry points. Model service access, official simulator environment/Assets and task results are separate from package integrity. See the release's LICENSE_NOTICE.md for the absence of an independent experience-data license statement.

LIBERO Long++, Ultra-Long and PRO v1.0.0

The three LIBERO releases preserve the exact sealed packages used by the delivered no-VLA/no-SAM3 evaluation images: 12 Long++ v12 packages, 9 Ultra-Long v12 packages, and 40 PRO persistent-memory v18 packages mapped to 80 Task/Swap cases. Each release contains a frozen case map, machine-readable package catalog and complete checksum manifest. Download the selected <benchmark-id>/v1.0.0/* path plus scripts/verify_package.py at a full repository commit, then verify it with the checksums_sha256 value in packages.json. Simulator assets, model weights, credentials and evaluation results are not included.

RoboCasa365 Target50 v1.0.0

The release preserves the exact task memory built into the delivered HEART RoboCasa evaluation image ghcr.io/opensensenova/sensenova-roborsi-robocasa365@sha256:8494fef17ee5979a072967632a2cb6ca885a165e87b01fb88d40c9f88b5b82e1: 111 files from RLinf/RPent-memory robocasa/results/ at revision 551fc3157b3e56b40a3d3a3b4c7ff81721ebe89b, split into 43 per-task packages. Each cell uses only its own task's seed-0 memory; 7 tasks have none. Download robocasa365/v1.0.0/* and scripts/verify_package.py at a full repository commit and verify with the checksums_sha256 in packages.json. See the RoboCasa usage guide for mounting the merged memory root into the image. Simulator assets, VLA weights, credentials and evaluation results are not included.

BEHAVIOR v1.0.0

See the BEHAVIOR release guide for 100 task packages with 22,935 frozen payload files. All 100 source ZIP inventories were verified; retained files are byte-identical, while 693 generated bytecode files are omitted. Original metadata is retained outside the evaluation payload. Download behavior/v1.0.0/* at a full repository commit and verify it with the checksum in packages.json. The updated RoboRSI downloader uses sealed task directories directly; it does not accept the old task ZIPs. These are existing ICL materials, not new simulator/model results.