dlsmarta/humanoid-vla-simple-token-conversions

Humanoid VLA: SIMPLE expert demonstrations → pure SONIC actions

This dataset converts official SIMPLE expert demonstrations into training data for a humanoid vision-language-action policy.

Training target:

image + language + proprio  →  VLA  →  action[T,128]
action[T,128] = body SONIC[T,64] ++ hand SONIC[T,64]

The 128-D action contains only SONIC tokens. Navigation commands, base height, torso velocity, turning flags, target yaw, teacher state, and future trajectories are not part of the VLA action.

Latest conversions

Current consumer entry point: simple/latest/manifest.json

It lists the newest conversion folder plus environment-success and semantic-success episode IDs for every task.

Three different training products

Every completed task exposes these separately:

1. All-demo labels cover the full official source denominator. They are source-aligned supervision and are not automatically closed-loop successes.
2. Executable success tokens are the exact frozen body64 ++ hand64 payloads, complete initialization state, and scenes that passed strict source-free replay. Use this product to reproduce successful unrolls.
3. Unified strict-success training roots project only refined semantic successes into self-contained LeRobot v2.1 roots with action, pre-action ego video, proprio, initialization, scene, quality/lineage, and consumer smoke together.

Success IDs apply to the executable/refined payloads. They must never be transferred onto different direct/best-effort action bytes merely because the episode ID matches.

Progress and success funnel

Every task separates environment and semantic success:

1. Source expert demonstrations — episodes provided by the official source corpus. The release has no per-episode success scalar, so 99 is a corpus count, not 99 independently verified task successes.
2. Native expert replay — demonstrations whose recorded WBC commands still complete the task in the pinned current runtime.
3. Environment success — the official simulator predicate fires during frozen source-free replay.
4. Semantic success — the intended task is completed under task-specific geometry, motion-quality, and visual checks.

Task	Source demos	Native replay	Environment success	Semantic success	Status
BetweenTables	99	29	22	18	Complete
Handover	100	68	41	1	Semantic audit complete
XMovePick	99	81	19	18	Complete
XMoveBendPick	100	Pending	Pending	Pending	Queued
BendPick	100	Pending	Pending	Pending	Queued
TabletopGrasp	100	Pending	Pending	Pending	Queued

Current episode IDs

Machine-readable counts, folders, and revisions live in simple/latest/manifest.json.

BetweenTables

• Environment successes — 22: 1, 4, 12, 16, 34, 35, 53, 54, 57, 66, 67, 70, 78, 80, 82, 84, 85, 87, 89, 92, 93, 95
• Semantic successes — 18: 1, 4, 12, 16, 35, 53, 57, 66, 67, 70, 78, 80, 82, 84, 85, 89, 92, 95
• Environment-only motion outliers: 54, 87

Handover

• Environment successes — 41: 0, 3, 4, 7, 8, 9, 17, 18, 22, 23, 25, 26, 27, 31, 32, 33, 34, 35, 37, 42, 47, 51, 55, 56, 57, 58, 59, 61, 62, 68, 72, 73, 74, 79, 82, 83, 84, 90, 94, 95, 96
• Semantic successes — 1: 31

XMovePick

• Environment successes — 19: 17, 20, 24, 30, 31, 32, 34, 45, 56, 59, 60, 68, 70, 71, 73, 78, 80, 85, 91
• Semantic successes — 18: 17, 20, 24, 30, 31, 32, 34, 45, 56, 59, 60, 68, 70, 71, 78, 80, 85, 91
• Environment-only state rejection: 73
• Environment set ID: 0b4e1c6bcc4f9e38912c4d86a5357f6c76e202687a61288c64fab36e4fff2023
• Semantic set ID: c1e9312af36fe29a7b61abe8434d366c37830e7f269b5ff7bd141befafcdbc6c
• Visual grasp confirmation is AI-assisted review, not human review.

Pending tasks

• XMoveBendPick: pending
• BendPick: pending
• TabletopGrasp: pending

BetweenTables latest current-best replayable all-99 package

• Complete replayable denominator: all_valid — 99 episodes
• Semantic/motion-qualified replayable successes: strict_success — 18 episodes
• Environment-success replay payloads: selected-environment22-v1 — 22 episodes
• Immutable all-99 package: current-best-all99-89354f326973-v1
• Replay implementation: PR #9 · pinned renderer/runtime c78ad6a4
• Exact replay/consumer guide: consumer/README.md · runtime-contract.json

Every episode—including failures—contains selected frozen body+hand tokens, complete initial body/hand/root state, trajectory-free scene, strict replay evidence, source binding, selected SONIC action/video/proprio, official source ego video, and normalized comparison. Success selectors label the successful payloads; they do not control whether an episode is replayable.

Handover latest current-best replayable all-100 package

• Complete replayable denominator: all_valid — 100 episodes
• Environment-success replayable subset: environment_success — 43 episodes
• Semantic/motion-qualified subset: strict_success — 1 episode
• Immutable package: current-best-all100-6aa1b5167a26-v1
• Replay implementation: PR #9 · pinned package head 6080734
• Manual download/unroll guide: replay/README.md

Every episode—including all 57 failures—is replayable from token zero and includes selected tokens, complete initialization, trajectory-free scene, strict rollout evidence, official source ego, exact selected SONIC ego, action/proprio, normalized comparison, and lineage. The selected namespace preserves all 41 previous successful payloads byte-for-byte and adds successful episodes 29 and 38.

XMovePick latest current-best replayable all-99 package

• Complete replayable denominator: all_valid — 99 episodes
• Environment-success replayable subset: environment_success — 19 episodes
• Semantic/motion-qualified subset: strict_success — 18 episodes
• Immutable package: current-best-all99-cdef218509d1-v1
• Replay implementation: PR #9 · pinned package head 6080734
• Manual download/unroll guide: replay/README.md

Every episode—including all 80 failures—is replayable from token zero and includes selected tokens, complete initialization, trajectory-free scene, strict rollout evidence, official source ego, exact selected SONIC ego, action/proprio, normalized comparison, and lineage. The selected namespace preserves all 19 previous successful payloads byte-for-byte. A fresh packaged success replay reproduced episode 17 with all 28 replay-visible arrays byte-identical.

Unified strict-semantic-success LeRobot roots

Immutable data revision: a2d07c185e5024e6133d6fd1ea1929a7c1b76f40

• BetweenTables: 18 episodes / 10,862 rows — unified-strict-semantic18-eae44bf3bf70-v1
• Handover: 1 episodes / 270 rows — unified-strict-semantic1-e0bef2705ecf-v1
• XMovePick: 18 episodes / 4,120 rows — unified-strict-semantic18-b4d174188925-v1

Each root is self-contained and passed full MP4 decode, exact action/proprio alignment, dataset construction, __getitem__, and a [50,128] Pi0.5 transform. The manually gated repository still denies anonymous payload reads.

Use this for VLA training

Each completed task has this stable shape:

simple/<task>/
├── README.md                         # task card and exact revisions
├── generations/<generation>/
│   ├── converted-tokenft-v2/
│   │   ├── data/chunk-000/episode_XXXXXX.parquet
│   │   ├── videos/chunk-000/ego_view/episode_XXXXXX.mp4
│   │   └── meta/{info.json,episodes.jsonl,tasks.jsonl}
│   ├── proprio-tokenft-v2/           # rollout-derived q_dev29 + gravity3 + row mapping
│   ├── hand-tokenft-v2/              # aligned hand64 targets
│   ├── episodes/                     # label receipts and provenance
│   ├── episode-quality.jsonl         # source/native/strict quality axes
│   ├── dataset-summary.json
│   ├── files.json                    # closed size + SHA-256 manifest
│   └── COMPLETE.json                 # receipt-last generation seal
├── evaluations/<evaluation>/         # strict source-free replay evidence
└── visual-confirmations/<bundle>/    # videos and hash-bound media receipts

For model training, point the loader at:

simple/<task>/generations/<generation>/converted-tokenft-v2

Important columns:

• observation.images.ego_view: egocentric RGB at 50 Hz.
• action: float32 [128], exactly body64 ++ hand64.
• action.sonic_tokens: identical to action.
• action.motion_token: body64 prefix for the existing SONIC loader.
• The final policy table contains no source.action_wbc; the training target is pure SONIC.
• source_index / source_timestamp: exact source alignment.

Compute normalization statistics for observation/proprioception fields. Keep FSQ action targets on their raw 1/16 lattice; do not normalize the token actions.

Decoders required to unroll the tokens

Training predicts the 128-D token action directly; it does not run a motion decoder. Execution splits each action into body64 and hand64 and uses two frozen decoders:

• Body: NVIDIA GEAR-SONIC recurrent G1 decoder, model_decoder.onnx, obs_dict [1,994] → action [1,29], SHA-256 c7241a123eaa36b5d64bad19540efde93cac1ad443bd4572fd12ca99898118ed.
• Hands: HandSONIC Dex3 decoder, token [B,64] → out [B,10,14]; use center frame 5. Graph SHA-256 7525734429dfbd66d7aecf6ee9f446a260d4829d38fcdee62a57c21964b90988; external-data SHA-256 dcd954ff38249955240c82a1ccafdb6adbbad39e81fe1b5f37124f06d3203bf3.

The body decoder also consumes causal live recurrent state. That feedback is required for execution and is not an extra demonstration action channel.

See DECODERS.md for exact contracts, hashes, runtime pseudocode, source locations, and the training-versus-execution boundary.

BetweenTables all-99 photoreal ego views — stable phase-synchronized selection

Every BetweenTables episode now has two aligned deliverables:

• a SONIC-only observation.images.ego_view video for VLA training;
• a source-versus-SONIC side-by-side diagnostic video with explicit STRICT SUCCESS or BEST HYBRID ATTEMPT | STRICT FAILURE labeling.

The left panel is the immutable official source ego stream. The right panel is the final selected SONIC-token-only rollout in the full SIMPLE Isaac/RTX task scene. MuJoCo remains authoritative for physics and task success. Phase synchronization recovered episodes 35, 80, 84, 85, 87, 89. Episodes 34, 93 are environment-only because final semantic placement fails; episodes 54, 87 are environment-only because realized motion is too jittery. Episode 4 is shown below because its basket placement is visually unmistakable.

Representative source | SONIC-only unroll comparison by task

BetweenTables — episode 4

Open BetweenTables comparison MP4

Handover — episode 31 environment and semantic success

Open Handover comparison MP4

XMovePick — episode 91 environment and semantic success

Open XMovePick comparison MP4

Each completed task gets one row-aligned MP4: official source on the left, exact SONIC-token-only unroll on the right. Pending tasks are not given placeholder media.

Immutable task pointers

All current revisions, folders, and episode IDs are in simple/latest/manifest.json.

Why official source demonstrations do not all replay successfully

The official release is canonical training data, not a deterministic physics-replay package. It provides videos, a 36-D action, body/hand joint observations, torso height, and episode boundaries, but no per-episode success/reward field and no complete physical-state trace. In particular, the processed schema omits measured root XY and base velocity and uses command-derived orientation fallbacks.

The release identifies LeRobot codebase_version: v2.1, but does not pin the recorder's SIMPLE Git SHA, WBC/controller SHA or checkpoint, MuJoCo build, or asset hashes. This repository therefore separates four denominators instead of calling every source episode a verified success:

1. 99 released source episodes — corpus count; source success is not labeled.
2. 36 captured source traces reporting success — capture-path outcome.
3. 29 source-controller replays accepted in the stricter pinned runtime — native replay outcome.
4. 20 task-stable SONIC-token placements — source-free token replay plus stable final placement; 18 additionally pass semantic motion qualification.

Remaining replay differences can come from unreleased recording-runtime differences, missing replay-critical state, contact sensitivity, and the lossy FSQ-token/controller conversion. SIMPLE's current target/container contact predicate also has an asymmetric operator-precedence branch; this dataset therefore applies stricter symmetric released-placement, final-stability, and motion-quality gates. Exact canonical replay would require the original recording environment or a release that includes exact source revisions, controller/checkpoint identity, asset hashes, and full robot/object state.