REPORT
First real-robot ref-motion tracking: real ≈ sim
The X2 tracked a 29.8 s all-natural mocap clip (5 m out → turn in place → walk back) on the gantry; operator verdict: “real and sim videos nearly identical.” Zero falls, zero guard trips, first field validation of the v5 unified deploy stack; the 25° tilt gate refused one bad-pose engage. Residual tracking gap localized to the model (arms 26.5°, stride shrink ×0.68), not the deploy path.
2026-08-20 · SBS video + real-robot log
REPORT
Basis-64: data diversity confirmed — 9/14, new champion
Same official recipe, data widened 16 → 64 clips overnight: core-16 exam 9/14 (all-time high), mean cadence error 0.151 (all-time best). jog_backward_start passes at exactly 1.00 — never passed before. Champion basis64_5000.
2026-07-08 · video + gate table
REPORT
Basis-16 official-recipe run — WASD skill library
16 foundation clips for keyboard driving (walk/jog start-loop-stop, backward, sideways, turns, idle). Official rewards + large batch. RUN COMPLETE: 8/14 at the 3000 finale (trajectory 3→6→5→7→6→8); forward-walk family + both turns locked at 1.00. Superseded by basis-64.
2026-07-08 · final mastery board
REPORT
Overfit ceiling: 0.071 m gap, ~75% fidelity
Single-clip overfit sets the ceiling on X2: anchor gap 0.071 m, no micro-stepping, eyeball-verified ~70–80% fidelity. Proves the stack works — the fleet's shortfall is training recipe, not the robot.
2026-07-07 · SBS video
REPORT
Three X2 embodiment boundaries: high-dynamic ceiling ~75–85%
From URDF + motor specs: 3° shoulder adduction, 20% shorter legs, 11.9 rad/s leg-motor speed cap vs 14.1 reference peak. Hard embodiment constraints — not training issues.
2026-07-05 · summary card
REPORT
X2 data pipeline: x2_native_v3 at official-grade smoothness
Canonical pipeline = GMR retarget + wrist fix + 6 Hz smoothing → 2,497 clips, jerk 7.0 ≈ NVIDIA official 6.9. All X2 training data now uses this pipeline. With a 7-clip sample gallery: G1-official ↔ X2 canonical + human → X2 retarget stage.
2026-07-04 · SBS video + gallery