---
license: unknown
language:
- en
tags:
- robotics
- quadrupedal-loco-manipulation
- simulation
- lerobot
size_categories:
- n<1K
configs:
- config_name: demonstration_index
  data_files:
  - split: index
    path: metadata/demonstration_index.parquet
---
# QLM-Bench

QLM-Bench provides quadrupedal loco-manipulation task environments, scripted data
creation and a physical interface for external high-level policies. The current
system uses a fixed Unitree Go2, the front-left manipulation leg, the original
RAMBO whole-body controller and native9 commands with zero desired force. The
three current tasks are **Push Box**, **Lift Basket** and **Press Button**.

This dataset repository currently contains **90 frozen usable scripted
demonstrations** in four independent releases, their Raw/Canonical data,
provenance and replay. [The catalogue](catalog.json) and
[episode discovery index](metadata/demonstration_index.json) point to those
existing releases. The index is metadata, not a combined training split or a
model-input table. Physical execution, resource loading, CPU data tools and
policy-independent evaluation interfaces are separate code responsibilities.

The initial catalogue metadata update is bound to [QLM source commit `b9be357048ae350dd006eef9c95b4c84b7ec4e45`](https://github.com/JalenLoong/QLM-Bench/tree/b9be357048ae350dd006eef9c95b4c84b7ec4e45) on the
[quadruped/native9 synthesis branch](https://github.com/JalenLoong/QLM-Bench/tree/synthesis/v2_9D-action_ego%2Btask-centric_camera).
The [frozen source/specification snapshots](metadata/spec_snapshots/qlm-core-0.1.0-b9be357048ae350dd006eef9c95b4c84b7ec4e45/snapshot_manifest.json)
record that committed source and content identity. Existing demonstration releases
remain unchanged; this update adds discovery metadata and declarative snapshots.

Local checks cover CPU fixed-data compatibility and bounded three-task Isaac
runtime lifecycle/regression paths. The public collection path has also completed
a 32-command diagnostic attempt with validated Raw/Canonical/media and an explicit
timeout; that attempt is local-only and is not an eligible new demonstration.
These checks do not establish whole-task success rates, a reconstruction-verified
benchmark suite, WAM learned-policy closed-loop inference or a formal score.
The initial catalogue-only update introduced no asset payload, case set, result release or new demonstration. The subsequent Normal20 release below adds20 usable demonstrations.

## Existing demonstration releases

| Task | Release | Episodes / actions | Standalone train / validation / test |
|---|---|---:|---|
| Push Box | [DATA-006-20260916T075451Z](releases/DATA-006-20260916T075451Z/manifest.json) | 50 / 24,335 | 40 / 5 / 5 |
| Lift Basket | [DATA-009-local-ten-20260926T130524Z](releases/DATA-009-local-ten-20260926T130524Z/manifest.json) | 10 / 7,042 | 8 / 1 / 1 |
| Press Button | [DATA-010-20260926T073438Z](releases/DATA-010-20260926T073438Z/manifest.json) | 10 / 5,194 | 8 / 1 / 1 |
| Push Box · Normal supplement | [QLM-DATA-001-20261008-normal-v2](releases/QLM-DATA-001-20261008-normal-v2/manifest.json) | 20 / 9,569 | 20 / 0 / 0 |

The historical entries in [the release index](releases.json), their release manifests, payload bytes,
episode/attempt identities, source mapping, timestamps and splits remain unchanged. New entries are appended.
Only usable demonstration Raw/Canonical and corresponding evidence/replay belong
to the historical releases. Pilots, failed/quarantined collection attempts, model
weights/caches and training artifacts remain excluded. This demonstration
eligibility rule does not filter future evaluation trial outcomes: a result release
must retain its complete predeclared trial table, including failures and exceptions.

Push Box instruction: **Push the box into the target area.** Success uses the
full XY envelope of all eight pose-transformed source-box corners inside that
 episode's goal, with the robot not fallen for three consecutive 50 Hz ticks.
Roll/pitch are part of the geometry; contact is diagnostic/unknown and is not a gate.
The original ten DATA-005 episodes and their 8/1/1 split remain present; forty
DATA-006 additions produce the existing 40/5/5 total. Original resource/scenario
variations and per-attempt source hashes stay in the release metadata.

Lift Basket instruction: **Lift the basket.** Success uses actual whole-basket
minimum clearance of at least 2 cm, not fallen, for three consecutive 50 Hz ticks.
Accepted demonstrations thread the complete FL foot before lifting. Only initial
XY varies in this ten-episode release. The first two episodes retain the accepted
visibility-gated expert; the remaining eight remove visibility blocking following
user clarification. Partial ego cropping remains disclosed and diagnostic. Old
6 cm pilots, other pilots and the interrupted attempt are excluded and are not relabeled.

Press Button instruction: **Press the button.** Success uses the accepted axial
button displacement threshold of 12 mm and a three-tick hold, with the robot not
fallen. Industrial and emergency-stop source assets, cap-color/scenario variations,
fixed cameras/controller and original task parameters are recorded per episode.
The existing release acceptance and source fingerprints remain the authority.

## Read an explicit release

Storage schema remains `wam-quadruped-v2.1.0`, LeRobot codebase v2.1 and the
original native9 action identity `rambo-native9-v2`. Raw contains N+1 boundaries;
Canonical contains N action rows and paired 50 Hz RGB frames plus separately saved
pre-reset terminal state/two RGB PNGs. Controller/physics rates are 100/500 Hz.
Observer RGB is 25 Hz and diagnostic-only. Timestamps are authoritative source
`simulation_time_ns`; they are not synthesized from frame indices.

Download one fixed revision and explicit Canonical root, including its release manifest:

```python
from huggingface_hub import snapshot_download
snapshot = snapshot_download(
    repo_id="dontKnow23456/QLM-Bench", repo_type="dataset",
    revision="4b5e0ed9aa04cbc1143774798f63171b27d838a7",
    allow_patterns=[
        "canonical/lerobot_v2_1/DATA-006-20260916T075451Z/**",
        "releases/DATA-006-20260916T075451Z/manifest.json",
    ],
)
```

Read data only from that Canonical root using its schema. Do not recursively scan
all Parquet files in this repository; the discovery table, future cases and future
results are separate resource types. The dedicated `demonstration_index` viewer
configuration contains episode references only. Original model-specific caches and
normalizers remain consumer-owned and outside Canonical.

For three-view Raw replay, download the selected full pinned release, serve the
snapshot directory with `python -m http.server`, and open that release's
`releases/<release>/review/index.html`. Historical replay paths remain portable
and unchanged. No old video or terminal image is re-encoded by this catalogue update.

## Push Box Normal20 supplement · 2026-10-12

[QLM-DATA-001-20261008-normal-v2](releases/QLM-DATA-001-20261008-normal-v2/manifest.json) adds20 successful Normal
demonstrations (slots `push_box__ep0081`–`push_box__ep0100`),9,569 completed
50Hz native9 actions, original dual-view50Hz RGB and25Hz observer videos.
Box01 and Box04 each contribute10 episodes; all use the frozen Randomized/ap_03
conditions. Continuous phase handoff removes the earlier abrupt FL target
retraction. Full-footprint/not-fallen/three-tick success, terminal-before-reset,
Raw/Canonical/media validation and initial object visibility checks passed.

This is a **train-only supplement:20 train,0 validation,0 test**, retaining
the original slot assignments. It does not complete the planned100-episode task,
provide the requested new Recovery quota or automatically create a combined
training/evaluation split. It is `publication_only` for the historical first
WAM v2/v3 runs; their original dataset revision, caches and normalizers are unchanged.
Only accepted trajectories are included. Failed/interrupted attempts, retry logs,
earlier unapproved Normal/Recovery data, assets and model payloads remain excluded.

Publication source: [QLM commit189545f](https://github.com/JalenLoong/QLM-Bench/commit/189545fbf0c00125f751f6f9bfb2c009f382319b),
branch `scaleup/v1/push_box`. Actual collection source commits and dirty-source
hashes remain preserved per episode; publication does not relabel their provenance.
Immutable payload revision: `87f877910effbae4d20427d9243296d0f9cd4a58`.

- [Canonical collection](canonical/lerobot_v2_1/QLM-DATA-001-20261008-normal-v2)
- [Raw data and all three camera views](raw/v2/push_box/QLM-DATA-001-20261008-normal-v2)
- [Portable three-view replay](releases/QLM-DATA-001-20261008-normal-v2/review/index.html)
- [Canonical replay and terminal frames](releases/QLM-DATA-001-20261008-normal-v2/review/canonical.html)

For local replay, download this release at the payload revision, serve the
snapshot root with a local HTTP server and open its review HTML. Video files
are retained without additional preview re-encoding.

## Historical WAM first-run input declaration

First WAM adaptation v2 **and v3** training, validation, test and evaluation remain
**Push Box only**: revision `4b5e0ed9aa04cbc1143774798f63171b27d838a7`, release
`DATA-006-20260916T075451Z`, exactly 50 episodes and the original 40/5/5 split.
The dataset manifest and Canonical inventory are pinned by content hash in the
[compatibility metadata](metadata/wam-first-adaptation-compatibility.json).
Existing model inputs and action/proprioception normalization remain unchanged.
WAM's historical 12.5 Hz RGB sampling and latent grouping remain WAM preprocessing
behavior, not a new benchmark requirement for every external policy.

Lift Basket and Press Button are `publication_only` for those first runs and are
excluded from their TRAIN/EVAL inputs. Their 8/1/1 splits do not expand that scope.
These are historical experiment-use declarations; they do not restrict future
independent use of the three task releases. There is no combined 70-episode WAM
first-run split. Integrity readback and scripted task success are not learned-policy evaluation.

## Resource and evaluation publication boundaries

Code owns TaskSpec, physical/runtime profiles, resource resolver and metric rules.
Large asset payloads require a complete dependency inventory, source/runtime hashes
and verified redistribution terms before being selected for a resource release.
No asset binary is included in this metadata update: Push Box source-license clarification,
OASIS redistribution terms/runtime shader closure and any remaining resource
publication checks are tracked separately. Button source declarations/attribution
also do not grant a license to the entire dataset. The existing RAMBO controller
remains an external, fixed third-party model dependency and is not mirrored here.

No reconstruction-verified formal EvalCase suite or official result release is
published by this metadata update. Source snapshots are generated from the committed QLM declarations and bound to
`b9be357048ae350dd006eef9c95b4c84b7ec4e45`; no formal evaluation protocol is silently added. Empty
cases/results or engineering test fixtures are not inserted as published resources.
Future results must declare protocol, suite/cases, assets/runtime/controller,
policy/code/checkpoint/adapter identities, expected trials, retry/error rules and
complete outcomes before a score is considered available. Horizon, seed count and
cross-task aggregation are not silently assigned by this catalogue.

## License, attribution and provenance

The dataset remains public with **`license: unknown`**. Data license clarification
is pending; public visibility is not an additional license grant. Source-code,
third-party asset, controller-model and demonstration-data licenses are distinct.
No new authors, paper acceptance or official benchmark release is claimed here.

Box source: [1057822006-svg/usd-cardboard-box-assets](https://github.com/1057822006-svg/usd-cardboard-box-assets),
release v0.1 at `cb386b4cba6c6adf2d42b54be37a78f9d4f01589`; its README recommends
CC BY 4.0 and requests attribution. The historical source manifest's pending
approval field retains its acquisition-time meaning. Basket source:
[TeleHuman/OASIS](https://github.com/TeleHuman/OASIS). Button sources are the
Industrial Button listing by Mark/@mark_saba (embedded USDZ credits Zeppelin) and
Emergency Stop Button by Miljan Bojovic/phoenix-storms; the official source
CC BY 4.0 declarations and both industrial-author attributions remain recorded.
Robot: Unitree Go2. Whole-body controller: RAMBO/CRL2 upstream.

Original release evidence, task acceptance, source/checkpoint hashes and camera
parameters remain under each release's immutable manifest/evidence paths. This
initial catalogue update changed discovery metadata, source/specification snapshots and the introductory card; it did not
rewrite the provenance of past runs or redistribute their third-party asset binaries.