Lesson 2 · 15 min
2026 whole-body policies learn from hours of human motion. Before using that data, you need to know how to read it and whether you are allowed to use it. In this lesson you open a BVH clip, count frames, fps and joints, write a clip of your own, and understand why the course does not redistribute any public dataset.
Start the lesson
Goal: read a motion capture clip (BVH) and report frames, fps and joints, knowing where the data comes from and under what license.
Click each block.
Motion data is training input: license and integrity come before quantity.
Click each one.
A collection of many captures converted to the same parametric body (SMPL-X). It is the most common input for GMR and for several motion tracking works.
Requires registration on the site. The terms for an open course were not checked: read them before using it and before publishing anything derived from it.
An animation dataset in BVH used in many retargeting examples (GMR has a ready-made configuration for it).
Same license caveat: not confirmed by the training's research on 10/19/2026.
The cleanest path for a course, a product or an article: capture your own motion (inertial suit, cameras, VR) or generate a synthetic clip.
The kit's bvh_minimo.py writes a BVH from scratch: a 5-joint skeleton and a wave. It is not a real human, but it shows the format and the license is yours.
The README of GR00T-WholeBodyControl cites the BONES-SEED dataset (03/16/2026, more than 142 thousand motions, about 288 h, with G1 trajectories in MuJoCo). The license was not checked: read it before using it.
The GMR repository has 1.2 GB in the assets folder, because of the robot meshes. The sparse-checkout downloads only the clip's folder, 2.9 MB (measured on 10/19/2026, commit bb1bbe4). ler_clipe.py uses no extra library: it reads the BVH text.
F=${FTH:-$HOME/fth}; cd $F
[ -d GMR ] || git clone -q --filter=blob:none --sparse https://github.com/YanjieZe/GMR
cd GMR && git sparse-checkout set assets/xsens_bvh_test
git checkout -q bb1bbe4 && git log -1 --format='GMR %h (%ad)' --date=short
cd $F/corpo-inteiro && . $F/.venv/bin/activate
python3 kit/ler_clipe.py $F/GMR/assets/xsens_bvh_test/251021_04_boxing_120Hz_cm_3DsMax.bvh; echo "saída: $?"HIERARCHY
ROOT Hips
{
OFFSET 0.000000 0.000000 0.000000
CHANNELS 6 Xposition Yposition Zposition Yrotation Xrotation Zrotation
JOINT Chest
{
OFFSET 0.000000 10.149404 -1.127712
CHANNELS 3 Yrotation Xrotation Zrotation
(…)
MOTION
Frames: 4249
Frame Time: 0.008333GMR bb1bbe4 (2026-04-02) arquivo: 251021_04_boxing_120Hz_cm_3DsMax.bvh raiz: Hips · juntas: 23 · pontas (End Site): 5 · canais por quadro: 72 ordem de rotação: Yrotation Xrotation Zrotation · unidade provável: cm (maior OFFSET 43.3) quadros: 4249 · fps: 120 · duração: 35.4 s linhas de movimento conferidas: 4249/4249 saída: 0
23 joints with 3 rotation channels, plus 3 extra position channels on the root: 72 numbers per frame. Frame Time of 0.008333 s = 120 frames per second.
arquivo: cortado.bvh raiz: Hips · juntas: 23 · pontas (End Site): 5 · canais por quadro: 72 ordem de rotação: Yrotation Xrotation Zrotation · unidade provável: cm (maior OFFSET 43.3) quadros: 4249 · fps: 120 · duração: 35.4 s linhas de movimento conferidas: 4000/4249 saída: 1
The header promises 4249 frames and the file has only 4000 lines (interrupted download, partial copy). The script exits with code 1 instead of letting retargeting run with the end missing.
gravado: ~/fth/movimentos/aceno.bvh (90 quadros a 30 Hz) quadros: 90 · fps: 30 · duração: 3.0 s
90 frames at 30 Hz = 3 s.
Five joints (hips, spine, head, right arm and right forearm), 60 frames at 30 Hz: the arm rises in 0.5 s and the forearm waves.
F=${FTH:-$HOME/fth}; cd $F/corpo-inteiro && . $F/.venv/bin/activate
python3 kit/bvh_minimo.py && python3 kit/ler_clipe.py $F/movimentos/aceno.bvh; echo "saída: $?"gravado: ~/fth/movimentos/aceno.bvh (60 quadros a 30 Hz) arquivo: aceno.bvh raiz: Hips · juntas: 5 · pontas (End Site): 2 · canais por quadro: 18 ordem de rotação: Zrotation Xrotation Yrotation · unidade provável: cm (maior OFFSET 30.0) quadros: 60 · fps: 30 · duração: 2.0 s linhas de movimento conferidas: 60/60 saída: 0
3 channels per joint + 3 for the root position = 18. The same reading works for any BVH.
Read it, check it, and decide whether you can use it.
Run the clip block and note the frames, fps and duration.
Done when: python3 kit/ler_clipe.py ~/fth/GMR/assets/xsens_bvh_test/251021_04_boxing_120Hz_cm_3DsMax.bvh | tail -1 → linhas de movimento conferidas: 4249/4249.
Generate the wave with 90 frames at 30 Hz and read it again.
Done when: python3 kit/bvh_minimo.py --quadros 90 && python3 kit/ler_clipe.py ~/fth/movimentos/aceno.bvh | grep quadros → quadros: 90 · fps: 30 · duração: 3.0 s.
For a clip you want to use (AMASS, LAFAN1 or another), write three lines in ~/fth/movimentos/licenca.md: Fonte, Licença (with the link) and Pode redistribuir.
Done when: grep -c -E "^(Fonte|Licença|Pode redistribuir):" ~/fth/movimentos/licenca.md → 3.
The explanation appears after your answer.
Question 1 of 3
The header says Frames: 4249 and Frame Time: 0.008333. How long is the clip?skill
Question 2 of 3
Why doesn't this course copy AMASS or LAFAN1 clips into the repository?
Question 3 of 3
ler_clipe.py says 4000/4249 lines checked. What do you do?
Choose an option. The explanation appears after you try.