SONIC-VLA-BonesSeed — prompt → motion-token dataset for Unitree G1

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 7 motions on a Unitree G1. It is the training input for the architecture-side VLA model wsagi/GR00T-N1.7-G1-SONIC-BonesSeed (one prompt-conditioned GR00T → SONIC token → balanced full-body motion).

Variant — BonesSeed: motions sourced from NVIDIA's GEAR-SONIC demos (rooted in the BONES-SEED corpus). A sibling SONIC-VLA-LAFAN dataset/model (LAFAN-sourced motions) is under development; they share the exact same prompt → token → WBC pipeline and differ only in the motion source.

🔗 Generation code: vitorcen/isaaclab-experience (scripts/gear_sonic_record.sh + gear_sonic/data/vla_token_recorder.py → scripts/sonic_vla_npz_to_lerobot.py)

🎬 Demo videos

The SONIC WBC tracking each of the 7 reference motions in Isaac Lab — the rollouts this dataset was recorded from.

squat	do a forward lunge
kick	jump on one leg
dance	dance the macarena
walk and turn around

Provenance / lineage

This dataset was generated in simulation, not scraped or hand-modified from an existing dataset. The lineage:

bones-studio/seed (142K human motions, ~288h, G1-retargeted)   ← what SONIC itself was trained on
        │  NVIDIA trains the SONIC WBC
nvidia/GEAR-SONIC  (released WBC checkpoint + demo_robot_filtered.pkl, 7 demo motions)
        │  we roll out the SONIC WBC tracking those 7 motions in Isaac Lab and record
        │  (motion_token, ego_rgb, proprioception, root_quat) each step  → convert to LeRobot v2.1
wsagi/SONIC-VLA-BonesSeed   (this dataset)

• Direct upstream: nvidia/GEAR-SONIC — both the WBC that emits the tokens and the 7 reference motions (demo_robot_filtered.pkl, SOMA-derived deploy demos).
• Distant upstream: bones-studio/seed — the human motion-capture corpus SONIC was trained on.

Contents

Format	LeRobot v2.1 (meta/info.json, meta/modality.json, parquet + mp4)
Episodes	7 (one per motion: dance, macarena, kick, lunge, squat, jump, walk)
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

dance · dance the macarena · kick · do a forward lunge · squat · jump on one leg · walk and turn around

⚠️ Scope

A derisk / proof-of-concept dataset: 7 motions, 1 episode each — meant to validate the prompt → token → WBC pipeline, not to train a general policy. No augmentation; new-motion coverage is out of scope.

License & attribution

Derived from the nvidia/GEAR-SONIC release; the reference motions and the token-generating controller are NVIDIA's. Use under the NVIDIA Open Model License terms of the upstream release. Please cite GEAR-SONIC and BONES-SEED.