Canonical pipeline = GMR retarget + wrist-order fix + 6 Hz smoothing → x2_native_v3: 2,497 clips, jerk 7.0 ≈ 6.9 (NVIDIA official G1 data). All X2 training data now flows through this pipeline.
Data-quality comparison: official data vs x2_native_v3 (after smoothing).
Passed visual review before adoption. Note: even the official dataset carries ~5 cm/s residual foot skate — that is the reference bar, not zero.
3 · Sample gallery — G1 official ↔ X2 canonical
Left = NVIDIA official G1 data, right = the same motion in the X2 dataset. Note the embodiment differences playing out: shorter legs (smaller stride), narrower shoulder range.
The GMR stage of the pipeline: human mocap skeleton (left) driving the X2 body (right). These renders are from the retarget stage itself, before the final smoothing pass.
forward jogwalk loop 360°sideways walk 45°backward jog
SBS clips: left = reference motion data, right = policy rollout (Isaac Lab, full physics) · yellow ball = reference anchor · all runs single RTX 3090