Datasets:

wsagi
/

SONIC-VLA-BonesSeed-V2

license: other
license_name: nvidia-open-model-license
license_link: >-
  https://developer.download.nvidia.com/licenses/nvidia-open-model-license-agreement-june-2024.pdf
source_datasets:
  - bones-studio/seed
task_categories:
  - robotics
tags:
  - robotics
  - humanoid
  - unitree-g1
  - whole-body-control
  - sonic
  - gr00t
  - lerobot
  - motion-tracking
  - vla
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/chunk-000/*.parquet

SONIC-VLA-BonesSeed-V2 — prompt → motion-token dataset with physical onset augmentation (Unitree G1)

A LeRobot v2.1 dataset pairing a language prompt + ego-view + proprioception with the 64-dim FSQ motion_token of the GEAR-SONIC whole-body controller. It is the training input for wsagi/GR00T-N1.7-G1-SONIC-BonesSeed-V2, whose headline is fixing cold-start onset (one-shot motions self-starting from a standing pose, no bootstrap).

What's new vs V1 (wsagi/SONIC-VLA-BonesSeed): V1 had 7 motions × 1 clean episode. V2 adds physical transition augmentation to cover the deployment "switch-prompt-from-any-state" distribution:

For each target motion, a [1 s of another motion's momentum] ++ [target onset] token sequence is replayed through the SONIC WBC (real physics) and re-recorded as (token, ego_rgb, proprio) — not spliced as raw frames (which would be physically inconsistent). This teaches the policy to initiate each motion from many incoming dynamic states, not just from a clean stand.
LAUNCH-filtered: only rollouts where the target motion actually initiated (not frozen / not fallen) are kept. All *_from_kick sources are dropped — kick's high foot-lift momentum contaminates the target onset and 1 s isn't enough to settle it out.
4 stand episodes added so the policy learns a natural idle.


Format	LeRobot v2.1 (parquet data + per-episode mp4 video + `meta/` schema)
Episodes / frames	54 / 12 630
fps	50
Prompts (8)	`dance`, `do a forward lunge`, `dance the macarena`, `kick`, `squat`, `stand`, `jump on one leg`, `walk and turn around`
Key features	`observation.images.ego_view` (480×640 h264), `observation.state` (43-d), `observation.projected_gravity`, `action.motion_token` (64-d FSQ)
Embodiment	`unitree_g1_sonic` (29-DoF G1); `meta/modality.json` maps state/action groups for GR00T

The augmentation tokens come from the GT token sequences in V1; the recorded (obs, token) pairs are produced by physics rollout, so every frame is a physically-valid WBC state. This is the "physically-correct" route (vs naive frame-splicing) — see scripts/sonic_build_aug_transitions.py and scripts/sonic_merge_aug_to_lerobot.sh.

🔗 Generation code: vitorcen/LeSONIC.

Load

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("wsagi/SONIC-VLA-BonesSeed-V2")
print(ds[0].keys())   # observation.images.ego_view, observation.state, action.motion_token, ...

License & sources

Reference motions derive from NVIDIA's GEAR-SONIC demo corpus (rooted in BONES-SEED). Released under the NVIDIA Open Model License (see link above). The SONIC WBC and GR00T base are NVIDIA's.