The bundled export is a kinematic reference, not an executable robot dance.
All A/B/C clips passed an independent model-based re-evaluation on 2026-09-05.
The checked-in data/pose_retarget_qa.json contains exact metrics, thresholds,
source/reference hashes, and calibration assumptions; provenance.json pins
the model, manufacturer velocity limits, geometry and generated files.
Each clip contains 1,299 frames at 40 Hz, retaining original frames 1–1299. Source frame 0 has a large discontinuity and is excluded only from the derived dance, not deleted from the recording. The source-time offset is 0.025 s and the last source timestamp is 32.475 s. Playback is 32.475 s with the final frame held for one sample; the first-to-last sample span is 32.450 s.
| Clip | Mean limb-direction error | 95th percentile | Maximum target-position error | Maximum sole penetration |
|---|---|---|---|---|
| A | 3.86° | 11.77° | 13.88 cm | 2.88 mm |
| B | 2.44° | 6.15° | 7.85 cm | 0 mm |
| C | 2.59° | 5.96° | 6.18 cm | 0 mm |
All samples remain inside the model's joint-position limits. Adjacent-frame joint speeds remain within the manufacturer's per-joint velocity limits, allowing only numerical tolerance (maximum ratio 1.00000000000002). The unchanged acceptance gates are mean direction ≤18°, 95th percentile ≤40°, target-position error ≤18 cm, sole penetration ≤25 mm, joint-limit violation ≤0.00001 rad and speed-limit ratio ≤1.00001. These are engineering study criteria, not certification or hardware-safety limits.
Human and robot proportions differ: the poses are adapted, not reproduced exactly. Units and source axes are explicit assumptions recorded in each clip's calibration report. The reset from the last frame to the first is a viewer loop boundary, not a validated continuous robot transition. Recentring and side-by-side offsets are presentation transforms, not stage formation.
A separate, bounded NVIDIA ProtoMotions/BONES-SEED tracking experiment used the full B reference for two simulated seconds, under gravity, contacts and finite actuator forces. It remained upright, but failed its 20 mm maximum contact penetration gate: the measured maximum was 58.43 mm. Recorded-state inspection identified hand/thigh and other self-contact, not merely ground penetration. This exploratory result does not validate A, C, longer tracking, balance recovery or the real robot. No collision masks or thresholds were relaxed to obtain a pass.
The experiment used ProtoMotions revision
607ca7a0bb92e261120bcab8d9f97f28b3130ffc and a checkpoint with SHA-256
a59baa3e04a951e5cf0b4cc68f24ebaafa9272714226618b99a5017dfc805b4c.
Neither the controller nor checkpoint is included in these graphics examples.
There is no robot SDK, DDS endpoint or actuator command path.
Retargeting ran on an RTX 4090 workstation. The constrained IK and temporal path search used its CPU; EGL reference rendering used its NVIDIA GPU. Full B/C solves took about 36.29/29.11 seconds. A needed multi-stage candidate and temporal-path research; those B/C timings do not describe A or a complete fresh reproduction run. No paid cloud resources were used.
Rendering, pose checks, and dynamic tracking are distinct tests. The viewer uses fixed robot offsets and parent-relative quaternion interpolation so rigid links remain connected between source samples. A successful screenshot or build cannot override a failed pose report or imply physical feasibility.