Datasets:

wsagi
/

SONIC-VLA-LAFAN

A small LeRobot v2.1 dataset that pairs a language prompt + ego-view + proprioception with the 64-dim FSQ motion_token of the GEAR-SONIC whole-body controller, for 8 hand-picked LAFAN windows (fight / run / dance) on a Unitree G1. It is the training input for the architecture-side flow3 VLA (one prompt-conditioned GR00T → SONIC token → balanced full-body motion).

Variant — LAFAN: the reference motions are hand-picked windows from the LAFAN1 locomotion/ dance corpus (Ubisoft), retargeted to the G1. Its sibling wsagi/SONIC-VLA-BonesSeed uses NVIDIA's GEAR-SONIC demo motions. Both share the exact same prompt → token → WBC pipeline and differ only in the motion source.

🔗 Generation code: vitorcen/isaaclab-experience (the LeSONIC subproject) — pipeline in doc/sonic_vla_principles.html.

🎥 Sample windows

A few of the recorded windows (the SONIC WBC tracking the reference motion the tokens come from):

Fight 2

Fight 5

Run 3 — run in a circle

Dance 2 — spin, step back, and clap

What makes the LAFAN variant different: a two-tier trackability gate

LAFAN clips are foreign choreography to the frozen SONIC WBC, so a clip can't simply be "converted to a dataset". Each candidate window must first pass a physical-validity screen: we let the frozen WBC physically track the reference and ask "did the robot actually fall?" (root height + torso tilt), not the release's strict per-frame safety envelope (which terminates on a single fast-strike frame even when the robot never falls). Of 10 hand-picked windows, only 2 pass the strict gate but 6 never fall — those 6 (plus 2 trackable prefixes) are what this dataset records. Tokens are recorded with the strict deviation terminations disabled so the full window is captured; this does not change the WBC's output, only whether the episode resets.

Provenance / lineage

Generated in simulation, not scraped. The lineage:

LAFAN1 (Ubisoft ubisoft-laforge-animation-dataset)   ← reference motions
        │  retarget to Unitree G1 (MimicKit), hand-pick clean windows by seconds
robot_filtered windows (fight / run / dance)
        │  PHYSICAL-VALIDITY screen: keep windows the frozen WBC tracks without falling
        │  roll out nvidia/GEAR-SONIC WBC tracking each window in Isaac Lab and record
        │  (motion_token, ego_rgb, proprioception, root_quat) each step → LeRobot v2.1
wsagi/SONIC-VLA-LAFAN   (this dataset)

Token producer: nvidia/GEAR-SONIC — the frozen WBC that emits the tokens (itself trained on bones-studio/seed).
Motion source: LAFAN1 — the Ubisoft locomotion/dance mocap corpus, retargeted to the G1.


Format	LeRobot v2.1 (`meta/info.json`, `meta/modality.json`, parquet + mp4)
Episodes	8 hand-picked windows (6 full + 2 trackable prefixes); 5,777 frames
Embodiment	`unitree_g1_sonic` (29-DoF G1)
fps	50
Key features	`action.motion_token` (64), `observation.state` (43), `observation.projected_gravity` (3), `observation.images.ego_view` (480×640), `annotation.human.task_description` (prompt)

The action.motion_token is the post-FSQ quantized token the SONIC actor emits at decode("g1_dyn", …) — i.e. the exact latent the controller consumes.

Prompts (8 windows)

fight · combat strikes and combo kicks · block and push-kick · fierce swings run · jog forward then run backward · sprint back and forth then backpedal · run in a circle dance · moonwalk · spin, step back, and clap

⚠️ Scope

A proof-of-concept dataset: 8 windows, 1 episode each — meant to demonstrate that LAFAN motions the frozen WBC can physically track can be absorbed into the same VLA token space (the "flow3" demo), not to train a general policy. No augmentation. "Physically valid" (never fell) ≠ "visually faithful" — fidelity is judged separately in closed-loop GUI.

License & attribution

The token-generating controller and its training corpus are NVIDIA's (nvidia/GEAR-SONIC); use under the NVIDIA Open Model License terms of the upstream release. The reference motions are from Ubisoft's LAFAN1 dataset — respect its license. Please cite GEAR-SONIC, BONES-SEED, and LAFAN1.