QLM-Bench

Quadrupedal loco-manipulation demonstrations generated with Go2, a single FL manipulation leg, the existing RAMBO controller, native9 commands and ego + task RGB. The dataset is public; the dataset license is pending (unknown). Public access is not an additional license grant.

Releases

Release	Task	Usable episodes	Actions	Train / validation / test
DATA-006-20260916T075451Z	Push Box	50	24,335	40 / 5 / 5

Payload revision: 21b69e2d775cd77d6565e25d1b0b99ad725b54b3. See release index, checksums, distribution and acceptance.

Only usable demonstration Raw/Canonical and associated evidence/replay are uploaded. Pilots, failed and quarantined attempts remain local and are not part of this repository's published payload. Existing release files are immutable; future releases use new directories. Model weights, asset source files, model caches and training artifacts are excluded.

Task and demonstration source

Instruction: Push the box into the target area.

Success requires the entire XY footprint envelope of all eight transformed source-box corners inside the episode's goal, with the robot not fallen, for three consecutive 50Hz policy ticks. Footprint calculation accounts for roll/pitch. Contact is diagnostic/unknown, not a success gate; unavailable bool placeholders must not be interpreted as confirmed non-contact. Robot spawn remains the native origin; desired FL force and external object wrench remain zero.

These are scripted high-level demonstrations, not manual teleoperation or learned-policy evaluation. The controller is unchanged. Ten original DATA-005 episodes retain their original8/1/1 split. Forty additional DATA-006 episodes add Box05/01/04, quarter-turn initial yaw, goal-border colors/geometry and smooth base/EEF command variations. Added splits32/4/4 give40/5/5. Durations are8.50–14.42s.

The50 episodes include Box05/01/04 counts30/10/10 and initial yaw0/90/180/270 counts20/10/10/10. Motion labels:10 baseline,14 coordinated,13 leg-finish and13 body-finish. Leg-finish is a configuration label;12 of13 actually triggered its conditional finish phase. demo_13-a4 and demo_25-a4 executed stop-forward, left body shift and inward FL corner nudging. Natural object tilt/toppling and residual yaw are reported rather than hidden; box upright/contact/yaw thresholds are not success gates. No generalization or trained-policy performance claim is made.

Contract and storage

Dataset schema wam-quadruped-v2.1.0; LeRobot codebase v2.1 / library0.3.3; action identity rambo-native9-v2. See contract.

Native9: base vx, base vy, base yaw rate, FL position xyz, desired FL force xyz. Command/state units and frames are specified by the included action config and contract. Raw/canonical policy RGB/actions are50Hz; observer25Hz is review-only. Model RGB12.5Hz uses factor-four row selection with inherited authoritative simulation_time_ns, never frame-index-generated timestamps. Canonical containsN action rows andN paired video frames with a separate pre-reset terminal state and two terminal PNGs. Raw retainsN+1 boundaries. All published episodes passed terminal-before-reset. Observer is never a canonical training camera or model input.

Canonical at canonical/lerobot_v2_1/DATA-006-20260916T075451Z is self-contained for model preprocessing. Model normalization is fitted only on the40 training episodes; VAE/T5 caches stay outside the dataset and are regenerated locally. Historical source manifests retain original paths and identities; the release manifest supplies portable Hub paths and source mapping. The release is a versioned bundle: do not treat every Parquet in this repository as one flat Hugging Face training table.

Download and replay

from huggingface_hub import snapshot_download
from pathlib import Path
snapshot = Path(snapshot_download(
    repo_id="dontKnow23456/QLM-Bench", repo_type="dataset",
    revision="21b69e2d775cd77d6565e25d1b0b99ad725b54b3",
    allow_patterns=["canonical/lerobot_v2_1/DATA-006-20260916T075451Z/*"],
))
canonical_root = snapshot / "canonical/lerobot_v2_1/DATA-006-20260916T075451Z"

Use explicit WAM/RAMBO resource paths with this canonical_root; keep model cache in a separate directory. Standard dataset-wide loading is not a substitute for the included native9 contract adapter.

For portable three-view replay, download the full pinned revision (omit allow_patterns), serve the downloaded directory with python -m http.server 8000, and open http://localhost:8000/releases/DATA-006-20260916T075451Z/review/index.html. The static replay files use relative paths to published Raw videos. The Hub may download HTML rather than render it directly.

Attribution and provenance

Box assets: 1057822006-svg/usd-cardboard-box-assets, releasev0.1, commitcb386b4cba6c6adf2d42b54be37a78f9d4f01589. The source README recommends CC BY4.0 and requests creator attribution. Its statement concerns source assets and is not a declaration of this dataset's license. See source attribution and historical source manifest. The latter's pending-approval field records acquisition-time history, not current task acceptance. Task approval and actual runtime geometry are preserved in per-episode metadata.

Robot: Unitree Go2; demonstrations use the existing RAMBO controller/checkpoint. Source/checkpoint hashes and camera/task parameters are retained per episode. RAMBO_Data owns simulation/publication; WAM-Policy owns model-side preprocessing/training. Model weights and third-party asset binaries are not redistributed here.