Shadow Dance v2.0.0 reference dataset

Team SELTZER's team-authored synthetic Unitree G1 reference dataset for Ultimate Bots Trial 03, upgraded from Shadow Dip to Shadow Dance: Dip + Gancho.

Evidence:

• 54 motions: 34 train/rehearsal, 8 selection-validation, 4 disclosed preflight-only legacy motions, 8 fresh final-test motions
• 54/54 manifest-bound reference checks pass, with 0 failures, 0 warnings, and 0 MuJoCo self-contacts
• All 30 v1 sequences and all 60 of their PKL/CSV hashes remain byte-identical
• Source commit: dc393f1
• Manifest SHA-256: 20803a03d9e3ddf3c7d381f59fb35fa83faf9e0dcb22e42f6dd25d51c7d21bb1
• Validation SHA-256: a9199a29d258d50eae2408bc55df3cb4989e2c5107734e684a777a15dbcec5b4
• Archive SHA-256: c1bde31a71e5d596f5018e01da8bcdb097ae314bd65c98c2aefa69fadb84217b
• Reference video SHA-256: 6a43b20311fe299a4687858d748e20ba20549dc72fd251398b9dc9bcd026f714

The four v1 policy probes conducted before v2 are explicitly isolated under preflight and excluded from training, selection, and final reporting. The eight v2 final-test inputs have not been supplied to a policy evaluator.

The MP4 is visibly labelled as a kinematic reference, not policy output. This release contains no fine-tuned checkpoint, derivative ONNX policy, official Isaac/WBT-Bench result, or stock/fine-tuned policy video; those remain unclaimed until licensed compute is run.

Motion Data by Bones Studio. No BONES-SEED motion or derivative is included.

Supplemental transparent stock preflight clip

• shadow-dance-v2-stock-mujoco-proxy-not-official.mp4
• SHA-256: 5d92810bf5313f85d55a5d07f93fffb477fee997be7c1cb739eb58253d94f6f1
• Exact committed heldout gancho target beside NVIDIA's public stock deployment model.
• On-screen labelled MuJoCo deployment proxy; not Isaac, WBT-Bench, a fine-tuned result, or a substitute for the required official before/after.

Experimental Shadow Dance affine proxy adapter - NOT official fine-tune

Experimental Shadow Dance affine proxy adapter — NOT an official fine-tune

This prerelease is transparent supplemental evidence for Team SELTZER's Ultimate Bots
Trial 03 entry. It is not the required Isaac Lab/PPO fine-tune, not a WBT-Bench
result, and not an official SuperSONIC score. It must not be represented as any of
those things.

The artifact appends a small supervised per-joint affine calibration to NVIDIA's public
GEAR-SONIC deployment decoder while retaining the public obs_dict [1,994] -> action [1,29] contract. It was fitted from 12 independently authored gancho training motions
(3,046 frames), selected on four separate validation motions, hash-frozen, and then
tested exactly once on all eight v2 test motions with no post-test retuning.

Complete frozen deterministic MuJoCo proxy result

Metric (macro over all 8 test motions)	Stock	Adapter	Change
Upright completion	8/8	8/8	unchanged
Local MPJPE	29.229 mm	29.448 mm	+0.75% (worse)
Global MPJPE	85.427 mm	84.810 mm	-0.72%
Root-position error	77.687 mm	76.706 mm	-1.26%
Joint RMSE	9.544 deg	8.628 deg	-9.60%; better on 8/8 motions

This is a mixed result: joint tracking improves, global/root errors improve slightly,
and local MPJPE becomes modestly worse. The proxy uses one deterministic MuJoCo 3.x +
CPU ONNX Runtime rollout per motion; it does not establish Isaac robustness or the
required three-seed evaluation.

Reproducibility anchors

• Source commit: ecf5c971c5d3351dbec2d01fe5161b501cdc9700
• Parent model: nvidia/GEAR-SONIC, revision
  9c0ff22b4ffec27c5392e8e284eb2f2df7a5b4e2
• Parent decoder SHA-256:
  c7241a123eaa36b5d64bad19540efde93cac1ad443bd4572fd12ca99898118ed
• Derived decoder SHA-256:
  caf8b40ca7de141498b3d1160654a2531ce45c149f6768d6cc0a2474bd5add27
• Complete package SHA-256:
  cc837b20e44aedcd5ded6a7a0146663fd5e0726a4ee3eb40b4f262d29a1d89c0
• Comparison video SHA-256:
  bb9a7f636ff1cba48008ebe02311bcc4d15bb2c5751087881f40d3d760e558a3
• 23.93-second honest-partial showcase
  (shadow-dance-submission-showcase-honest-partial-v2.mp4) SHA-256:
  41b2d0ed7f3b5fddfd07eddd99907e75d94b388fdeef99ddcc4db3f2abb8bc5a
• Dip kinematic-reference source (visibly marked not policy output) SHA-256:
  3f2c6b68b4b592af7855aeacf953a84f08ec3bfac20848a7380578519e040d4e
• Frozen selection record SHA-256:
  508c3dd0c3eea175c91e227030745df5d14c1404abf1930cc83fbd8ceb724ca8

The package contains the derivative decoder, unchanged public encoder, full model card,
license and notice, frozen adapter and selection records, validation report, complete
stock/adapter test evidence, and a 51-entry SHA-256 manifest. The release also provides
the comparison/video manifests separately for easy review.

The independently authored v2 motion dataset and generator are available in the
shadow-dance-v2.0.0 release.

Safety and license

Simulation only; no real-robot deployment or safety claim is made.

The parent model and this derivative are distributed under the NVIDIA Open Model License
included in the package. Licensed by NVIDIA Corporation under the NVIDIA Open Model
License.

Challenge acknowledgement: Motion Data by Bones Studio. No BONES-SEED motion or
derivative is included in or used by this independently authored dataset.

Shadow Dip v1.0.0 reference dataset

Team-authored synthetic G1 reference dataset for the Ultimate Bots Trial 03 SuperSONIC challenge.

Evidence:

• 30 motions: 22 train/rehearsal, 4 selection-validation, 4 untouched final-test
• 30/30 reference validation passes, 0 failures, 0 warnings, 0 self-contacts
• Public Ubuntu/Python 3.11 regeneration and cross-platform semantic comparison passed
• Source commit: 684c6e8
• Manifest SHA-256: 1b2045380e09e6276c5ac4ff4c2bb1c7bd5903a974940f9928d7351b5f90a5d1
• Validation SHA-256: 5aedeedee8d775c34c0b5c67f235591829da281cfa9385b8b8a15b8c10a6b999
• Archive SHA-256: 94099f031b8a0b5ea36c809e705f77088342a6b54d73f9735508b146841c1370
• Reference video SHA-256: d9f6f4284e5cecbc80349d050786b2c876a26f1a93dd4ba6e3da8f9149efe0c3

The MP4 is a labelled kinematic target reference, not policy output. This release contains no trained checkpoint, ONNX policy, stock/fine-tuned simulator result, or BONES-SEED motion/derivative. Those are not claimed.

Motion Data by Bones Studio.