Robo-Xperience10M-Scores

Per-episode motion/quality scores for the Xperience-10M egocentric corpus, used for data curation when training SONIC-token VLAs. Generated 2026-08-13 by score_episodes.py (humanoid-vla repo, data/xperience10m_converter/) from the GMR retargets (motion.pkl), the corrected proprio sidecars (Robo-Xperience10M-Proprio v2, fixed gravity), and each episode's valid.npy occlusion mask.

Why

Uniform sampling wastes the training budget: 72.7% of all 1 s windows are idle (no locomotion, no arm activity), 5.4% are locomotion-active, 3.9% manipulation-active, 17.9% in between (corpus_summary.json has the full statistics). Weighting samples away from idle gives ~3.7x effective exposure to useful content at the same draw budget and removes millions of "human stands still" targets (a stand-still-attractor risk for the policy).

Files

• <uuid>/<ep>/scores.json — RAW per-window statistics (thresholds are applied downstream; re-tuning never requires re-scoring). One entry per 1 s (50-frame @ 50 Hz) window:
  • loco_speed (m/s): mean horizontal root speed, from GMR root_pos at source fps. Note: GMR translation is scaled to the G1 robot (~0.75x human scale).
  • leg_vel, arm_vel (rad/s): mean |joint velocity| over the 12 leg / 14 arm joints.
  • valid_frac: mean of the occlusion-validity mask over the window.
  • grav_z: mean |gravity_body z| (1.0 = upright; low = bad retarget/SLAM).
  • episode block: aggregates + root_jitter (mean |root accel|) + provisional class fractions (idle_frac, loco_frac, manip_frac) at the reference thresholds recorded in ref_thresholds.
  • main_task: the episode's caption string, for language-based filtering.
• corpus_summary.json — corpus-level aggregates over all 12492 episodes / 3.47M windows.
• window_weights_v1.npz — compiled sampling weights consumed by the trainer. Key "<uuid>/<ep>" -> float16 array, one weight per window. v1 policy: idle windows (loco_speed < 0.15 m/s AND arm_vel < 0.30 rad/s) -> 0.0 (hard drop); 40 episodes with grav_z_mean < 0.85 -> all zeros; everything else -> 1.0. Mean weight 0.271. Recompile with different thresholds in seconds: compile_window_weights.py --idle-w ....

Threshold validation

Classes were checked two ways: caption semantics (walk/carry-captioned episodes score 3.9x higher loco_frac than sort/wipe/fold ones; manipulation-captioned 2.7x higher manip_frac) and against standard activity-recognition boundaries (stationary cutoff 0.1-0.2 m/s; active multi-joint arm motion >~0.5 rad/s). Distributions are monotone- decaying (not bimodal), so the cutoffs are conventions, not cluster boundaries.

Trainer integration

SonicTokenDataset (humanoid-vla) samples window start times proportionally to these weights when SONIC_XPERIENCE_SCORES=<this repo's local path> is set; unset = uniform sampling, byte-identical to the pre-curation behavior. Zero-total-weight episodes are dropped (resampled), never uniform-fallback. Enabling curation changes the sampled state distribution — recompute norm stats for any config that turns it on.