dlsmarta/humanoid-vla-token-conversions

BetweenTables successful frozen token conversions

This dataset contains 14 accepted frozen token conversions from G1WholebodyLocomotionPickBetweenTablesTeleop-v0: 12 from the immutable original fixed protocol, one raw-archive phase-aware recovery, and one separately namespaced filtered-archive recovery.

• Current native replay admissions from the 100 raw teleop episodes: 26/100
• Original strict frozen source-free replay successes: 12/26 replay-admitted (12/100 raw episodes overall)
• Phase-aware successor recoveries: 1 — episode 84, strict replay success tick 720
• Published accepted episode IDs: 22, 23, 32, 36, 39, 40, 53, 73, 76, 79, 82, 84, 93
• Filtered-archive successor: official-filtered-a104/episode_000071, strict replay success tick 585, explicit source criterion 0.2

Raw and filtered numeric episode IDs are different identity spaces. The filtered successor is stored under successors/filtered-a104/; it is not relabeled as raw episode 71.

Each episode directory contains the frozen body/hand token payload, its receipt, conversion summary, strict-replay summary, strict-replay rollout telemetry and receipt, the pre-import isolation receipt, and the final episode summary.

VLA training-label generation

training-labels/filtered-a104-command-inversion-b30e2a9/ contains aligned labels for all 99 official success-filtered demonstrations: 60,284 rows total.

Each row contains 133 action values:

• 64 body SONIC tokens;
• 64 hand SONIC tokens;
• 5 causal command fields: base height, torso X/Y velocity, turning flag, and target yaw.

The generation contains 36 labels derived from WBC traces that succeed under the current reconstructed runtime and 63 explicitly marked best-effort labels from non-successful current-runtime traces. This split does not change the official source status: all 99 demonstrations came from SIMPLE's success-filtered release.

These labels are post-hoc command inversions, not authentic demonstration-time SONIC tokens. Plain stock-token replay was 0/99 successful, so token-only training should not silently discard the five external command fields.

Claim boundary

These are post-hoc conversions, not authentic demonstration-time SONIC tokens and not VLA rollouts. Body tokens were selected by offline task-objective search. Hand tokens remain frozen and open-loop during replay. Strict replay receives no source trajectory and performs no online token correction. Stock frozen decoders and the ordinary live low-level controller consume current robot state.

The official task predicate is authoritative. Exact token replay proves deterministic reproduction of the selected behavior; it does not establish generalization to another reset, scene, episode, or task.

Source-dataset correction

The 100 inputs came from the official simple-teleop/ raw teleoperation archive, not the success-filtered simple/ training archive. The official HF revision contains 100 raw teleop episodes and 99 replay-filtered training episodes; 99/100 raw final object states are geometrically within the task's 12 cm target/container XY region. Therefore, 26/100 is the admission rate of this project's current no-state-copy native replay under its runtime and 0.40 dwell criterion—not the success rate of the human demonstration dataset. See SCIENTIFIC_CORRECTION.md.

The original report and episode_outcomes.csv remain the immutable fixed-protocol accounting. See SUCCESSOR_RECOVERIES.md for episode 84.