QLM-Bench · MuJoCo1000

main contains 1000 user-accepted MuJoCo demonstrations, ten tasks ×100. The dataset remains private. The complete original Isaac main remains on Isaac-based at 519725c8bc77b35fd5d3b4ac8555a3e96a8876a7. No Isaac episodes are mixed into this MuJoCo main.

This release adds the accepted600 to the previous400. The old400 receive only user-requested prompt annotations and their necessary checksum/index dependencies; all their video, numerical Parquet, snapshot, command and terminal payloads remain byte-identical. Press Button's wording and entire per-task payload are unchanged. The previous400 main is retained in Git history at a8eafc5b1e6019e4431d0417d9064f9fd658cc03.

Task	Episodes	Exact current prompt	Release
push_box	100	Push the box inside the outlined area.	mujoco-push-100-20261015-v1
press_button	100	Press the button.	mujoco-press-100-20261015-v1
lift_basket	100	Lift the basket by its handle.	mujoco-lift-100-20261015-v1
kick_ball_to_goal	100	Kick the ball through the goal.	mujoco-kick-100-20261015-v1
sweep_block_to_zone	100	Sweep the block into the outlined area.	mujoco-sweep-100-20261017-v1
close_fridge_door	100	Push against the fridge door until the door closes.	mujoco-fridge-100-20261017-v1
close_oven_door	100	Push the oven door until the door closes.	mujoco-oven-100-20261017-v1
turn_valve	100	Rotate the valve lever upward to the upright position.	mujoco-valve-100-20261017-v1
pull_open_drawer	100	Slide your foot into the drawer handle opening and pull the drawer outward.	mujoco-drawer-100-20261017-v1
open_room_door	100	Press down the lever handle, then push the door open.	mujoco-door-100-20261017-v1

Every task has 80 train /10 validation /10 test and 20 Clean /80 Randomized. Totals are 800 train /100 validation /100 test, 200 Clean /800 Randomized, 964,151 actions /965,151 state boundaries. Concrete configurations differ across splits while sharing the approved ID support. There is no extra OOD track, clutter quota or forced Recovery quota.

Catalogue · 1000-episode index · Current prompts · Release summary · Local replay. Viewer rows are episode discovery metadata; actual time-step data are in the linked Raw and Canonical payloads.

Physical data and language annotation

Fixed Unitree Go2, front-left manipulation, original RAMBO/native9 with force-zero, physics500Hz /controller100Hz /action-state50Hz. Ego/Task RGB is640×360@50FPS; Observer is1280×720@25FPS and is never a policy input. Raw hasN+1 boundaries and independent pre-reset terminal; Canonical hasN action-aligned frames and explicit terminal state/images. All actions, state, contacts, clocks, splits, task physics and success predicates remain unchanged by this language revision.

The original400 preserve their actual mixed CUDA-MJWarp/OSMesa and NativeCPU/EGL source identities. The new600 are NativeCPU/original RAMBOCPU/NativeEGL. All six new tasks passed complete engineering checks and final per-task random10 user review. Drawer contains31 preserved nominal-source successes and69 approved scope-v5 successes; its10 original failed attempts remain historical evidence and are not released as successful demonstrations. Door uses real historical geometric bolt/socket exit, then>=15deg and original three-tick/not-fallen success. Old fixed20mm logic is not used. No success threshold, action schema, camera intrinsics or numerical trajectory changes.

Fridge/Oven/Drawer/Room Door Observer videos use the user-requested left-rear view, rendered solely from the same existing real snapshots. Their Ego/Task videos and physics were not rerun. Other tasks' Observer views remain unchanged.

user-prompts-20261017-v1 is an explicit annotation revision: current Raw task_text, Canonical task/action-segment text and discovery metadata use the exact prompts above. Original acquisition instruction/source manifests and historical user/engineering receipts remain provenance; they are not rewritten as new executions. The local text embeddings were regenerated with the pinned frozen text encoder, without gradients or optimizer steps. Model weights/caches are not dataset payload.

Portable layout

• canonical/lerobot_v2_1/<release>/: independent100-episode collections.
• raw/v2/<task>/<release>/<uid>/: original required high-rate streams, all camera videos, commands/capture/provenance and terminal records, with current language metadata.
• physical/v1/<task>/<release>/<uid>/: original50Hz full snapshots, command and terminal originals plus a portable bundle.json. The source trace manifest's local optional telemetry paths are historical; required high-rate data are in Raw.
• releases/: per-task manifests and exact engineering/user acceptance evidence.
• schemas/: unchanged Raw/Canonical/action/camera contracts.

Catalogue DataRefs pin payload revision b84488eca8a723767cef27fa455ec2ea8169f81c. Download final main for the catalogue/Viewer/replay index, then use the published relative paths; no producer workspace path or asset installation is required for the data reader.

The MuJoCo1000 demonstration release is data-only; separately versioned evaluation resources are described below. No WAM training, normalizer refit, repeated task-consumer validation, AMD execution, learned-policy success claim, or historical TRAIN/EVAL input change is implied. The demonstration release excludes policy weights/caches, failed candidates and development references. The separate evaluation runtime packs preserve their approved mesh/texture dependencies. License remains unknown: no blanket asset-source redistribution or MIT grant is asserted. The original OASIS basket provenance and unresolved independent asset license status are preserved; no asset source package is included.

Publication work: QLM-DATA-041. The current GitHub code commits are reported separately; historical physical runs retain their actual original source identities.

MuJoCo evaluation runtime resources — QLM-EVAL-001

benchmarks/mujoco-standard-test100-v1/manifest.json freezes ten held-out test configurations per task (2Clean8Randomized), original environment seeds, policy seed42, original task horizons, exact resource identities, and actual zero-step NativeCPU/RAMBOCPU/EGL reconstruction evidence. Formal policy trials and benchmark SR have NOT_RUN status.

assets/<asset_id>/<resource_identity>/runtime/ contains each complete original MuJoCo pack, with an outer manifest and private-dataset redistribution evidence. The three-file original RAMBO controller is also available at assets/rambo-go2-controller/iteration-2000/runtime/, copied byte-for-byte from dontKnow23456/rambo-go2-policies@65fa7b035ad16af78c2af2d64a3ac4aed9c81fdc. Its model_2000.pt SHA256 is 1cc5f68fe15e37ccabae26060d79a26a8c078ed465a81f6f729c2009b67ca706; the independent model repo is retained.

The benchmark resource registry binds exact paths/manifests. Source archives and source manifests supply the local code commits without a GitHub push. With the QLM hub extra installed, qlm evaluate fetch-case --revision FIXED_REVISION --case-id CASE_ID --output NEW_ROOT acquires suite metadata and only that case's scene/controller resources; no training demonstrations or policy weights are downloaded. WAM supplies its own checkpoint, original normalizers, prompt embeddings and frozen encoders.

This additive resource publication preserves all original training payload bytes, paths, catalog/DataRefs, splits and exact prompts. Dataset visibility remains private; existing fixed revisions and Isaac-based remain readable. User-confirmed redistribution is scoped to this private dataset, preserves original attribution, and asserts no blanket new SPDX license.

The separately retained local engineering evidence uses original Isaac50 v2/v3 step1000 policies for22 bounded4s episodes. It is not a MuJoCo SFT evaluation or benchmark score. No training, AMD execution or WAM weight publication accompanies this resource release.

Accepted evaluation observer source — QLM-EVAL-003

The active QLM source snapshot is 8456b9ab352445028340315627ed0a40087cf421, available through benchmarks/mujoco-standard-test100-v1/source/source_manifest.json and its exact archive/hash. Protocol and resource-registry snapshot source identities are bound to the same committed source tree. The prior execution source remains separately indexed for historical engineering evidence.

Every evaluation task uses the accepted observer view: left-rear and slightly elevated, with eye [-0.70,1.55,0.85]m and look-at [0.10,0,-0.10]m relative to base translation/heading,35deg vertical FoV and stable world-up. This changes only diagnostic GL cameras; both policy views, physics/controller and task predicates remain unchanged. The original scene packs are not rewritten.

All100 case files, expected trial slots/seeds/horizons, original reset evidence, runtime/controller resources and training objects remain unchanged. validation/observer-source-update-20261018.json records the narrow source change and preserved historical provenance. The source guard still rejects stale or altered runtime code. No additional episode rendering, inference, physics, formal benchmark or SR is introduced by this source-only update. Local development episode/cache payloads were evicted after validation; compact original outcome/provenance receipts remain historical.

Corrective segments — push_box r1 (sft-v1/QLM-DATA-001)

mujoco-push-corrective-r1-20261111-v1 is a separate, additive release of 107 corrective segments for push_box (92 train / 15 validation, 18,143 actions). It is not part of the 1000 user-accepted demonstrations above and is not listed in their catalogue entries, Viewer splits or demonstration index; its discovery rows are in metadata/corrective_segment_index.parquet.

Each segment starts at a takeover boundary inside a failed rollout of a learned student policy on a push100 train scene (mujoco-push-100-20261015-v1, scenes 001–080 only; no validation/test scene is used) and ends at the original success terminal. Actions come from a frozen scripted recovery routine (push_box_recovery-v1); segments were checked by engineering gates, not reviewed one by one by a user. The clock of each segment starts at 0 on its takeover boundary; the absolute tick in the source rollout, the takeover point (A, B or A=B) and the takeover-state hash are recorded per row. Validation holds out whole scenes (8 scenes). Physics is NativeCPU with the original RAMBO model_2000 on CPU; rendering is an offline state replay with OSMesa, Ego/Task 640×360@50FPS and Observer 1280×720@25FPS (diagnostic only).

Layout is the same as above: canonical/lerobot_v2_1/mujoco-push-corrective-r1-20261111-v1/, raw/v2/push_box/mujoco-push-corrective-r1-20261111-v1/<uid>/, physical/v1/push_box/mujoco-push-corrective-r1-20261111-v1/<uid>/ and releases/mujoco-push-corrective-r1-20261111-v1/manifest.json with engineering evidence. The student's pre-takeover half of each rollout is provenance only and is not published. Catalogue key corrective_segments pins payload revision b4d7ba6d6a21430d7a9fd087e5c9e38c22785e48. License remains unknown, as for the rest of this dataset. Intended use: the sft-v1/ALG-002 task-specific SFT together with push100 train; no training result is implied by this data release.

Corrective segments — sweep_block_to_zone r1 (sft-v1/QLM-DATA-002)

mujoco-sweep-corrective-r1-20261112-v1 is a second separate, additive release of 80 corrective segments, for sweep_block_to_zone (72 train / 8 validation, 52,219 actions). Like the push_box release above, it is not part of the 1000 user-accepted demonstrations, their catalogue entries, Viewer splits or demonstration index; its rows are appended to metadata/corrective_segment_index.parquet.

Each segment starts at takeover point B (the student has lowered the foot near the push face, about 2–7 cm on the -x side of its centre, before touching the block) inside a failed rollout of a learned student policy on a sweep100 train scene (mujoco-sweep-100-20261017-v1, scenes 001–080, one segment per scene; no validation/test scene is used) and ends at the original success terminal. Actions come from a frozen scripted recovery routine (sweep_recovery-v1); every segment passed engineering checks and a sample of segments was reviewed and approved by a user. Its lateral sweep is faster than the original demonstrations and turns the block more (final yaw change −31° to −3°, median −13°). The clock of each segment starts at 0 on its takeover boundary; the absolute source tick and takeover-state hash are recorded per row. Validation holds out whole scenes (8 scenes). Physics is NativeCPU with the original RAMBO model_2000 on CPU; rendering is an offline state replay with OSMesa (the original sweep demonstrations used EGL). The Ego view does not see the block, as in the original demonstrations.

Layout as above: canonical/lerobot_v2_1/mujoco-sweep-corrective-r1-20261112-v1/, raw/v2/sweep_block_to_zone/mujoco-sweep-corrective-r1-20261112-v1/<uid>/, physical/v1/sweep_block_to_zone/mujoco-sweep-corrective-r1-20261112-v1/<uid>/ and releases/mujoco-sweep-corrective-r1-20261112-v1/manifest.json with engineering and review evidence. Pre-takeover student halves are not published. Its corrective_segments entry pins payload revision b8e6ac02f6d0055441b11d031326550a0b7f9a2d. License remains unknown. Intended use: the sft-v1/ALG-004 task-specific SFT together with sweep100 train; no training result is implied by this data release.