SIMPLE BetweenTables successful SONIC tokens
This dataset contains 14 frozen token-only conversions that passed fresh source-free SIMPLE replay for G1WholebodyLocomotionPickBetweenTablesTeleop-v0.
VLA target
Use sonic_action_labels with shape [T, 128]:
- columns
0:64: body SONIC tokens; - columns
64:128: hand SONIC tokens.
No base-height, torso-velocity, turning, target-yaw, navigation, source-trajectory, or encoder-window fields are included in the action label.
sonic_token_codes contains the equivalent signed FSQ codes in [-16, 15]. Recover float tokens with codes.astype(float32) / 16.
Decode contract
Split each 128-D row into body64 and hand64. The stock recurrent body decoder still requires its ordinary live proprioceptive/action history. The hand decoder is open-loop. Initial state arrays are included for deterministic replay initialization; they are not VLA action targets.
pi0.5 training-ready corpus
pi05-training/tokenft-v1/ contains the complete verified corpus consumed by pi05_sonic_bhs3_between_tables_ft:
- LeRobot-style body-token Parquet files;
- retimed ego RGB MP4s;
- per-episode prompts;
- proprioception and hand-token sidecars;
- exact hand-decoder runtime asset for the bhs3 state;
- deterministic 12-train / 2-eval split manifest;
- a real episode-22 input preview and tensor bundle.
The VLA receives one RGB frame, the prompt, and a 110-D current state. It predicts a 50×128 chunk of body+hand SONIC tokens. See pi05-training/tokenft-v1/README.md.
Evidence
- Accepted episodes: 14
- Aligned token rows: 8,833
- Strict source-free replay success: 14/14
- Byte-exact replay arrays: 14/14
- Falls: 0
Numeric episode IDs from raw-teleop and official-filtered-a104 are separate identity namespaces.
- Downloads last month
- 19