E1 · GMR environment
10 minRun the install block and confirm the versions.
Done when: ~/fth/.venv-gmr/bin/python -c "import general_motion_retargeting" 2>/dev/null; echo "saída: $?" → saída: 0.
Lesson 3 · 15 min
Retargeting: from the human skeleton to the robot
A person is 1.70 m tall and has joints the robot does not; the G1 is 1.32 m tall and, in the Menagerie model, has 29 joints with limits. Retargeting is the computation that takes one to the other, frame by frame. In this lesson you install GMR in its own environment, take 10 s of boxing to the G1, and check every frame against the joint limits.
Start the lesson
In this lesson you will
- 1What retargeting solves: scale, different joints, and limits
- 2One environment per tool: the .venv-gmr with pinned versions
- 3Per-frame IK with mink: human points become robot targets
- 4Why the limit depends on the robot file
Goal: take a human clip to the G1 with GMR in its own environment and check that no frame goes past the joint limits.
01One retargeting frame
Click each block. GMR repeats this for every frame of the clip.
Click each block of the diagram to see what it does.
GMR runs on the CPU, in real time according to the paper (ICRA 2026); here, 10 s of clip in a few seconds.
02Why a separate environment just for GMR
Click each point.
1. torch
The GMR package imports PyTorch as soon as it loads (the kinematic model uses tensors), even for BVH retargeting. Putting torch in the .venv from the MuJoCo practices would make that environment heavy for no reason.
Solution: a .venv-gmr just for it, with the CPU torch from the official PyTorch index.
2. setup.py
At commit bb1bbe4 (04/02/2026), setup.py asks for, among others, proxqp, smplx and redis, but BVH retargeting uses the daqp solver (the default in the code) and does not need the others.
When code and the dependency list disagree, the code wins: the course installs GMR with --no-deps and pins by hand only what runs.
Program rule
One environment per tool and versions pinned per project. BeyondMimic pins Isaac Lab 2.1.0; mjlab uses MuJoCo Warp with uv; in a single environment, one breaks the other.
3. Limits
The G1 model that GMR uses (g1_mocap_29dof.xml) has the same joint order as Menagerie, but tighter limits on 15 of the 29 joints (measured on 10/19/2026).
Because the retargeting respects GMR's limits, the result fits within Menagerie's. The reverse wouldn't hold: always check against the model that will run it.
03Install GMR in .venv-gmr
Adds only the G1's code and model to the GMR clone (sparse-checkout add), creates the second environment and pins the tested versions. First time: the whole environment took 1.2 GB on a fresh bench, almost all of it the CPU build of torch (measured on 10/19/2026).
Terminal · GMR environment
F=${FTH:-$HOME/fth}; cd $F/GMR
git sparse-checkout add general_motion_retargeting scripts assets/unitree_g1
cd $F && [ -d .venv-gmr ] || python3.12 -m venv .venv-gmr
. .venv-gmr/bin/activate
pip install -q --index-url https://download.pytorch.org/whl/cpu torch==2.14.1
pip install -q mujoco==3.16.0 numpy==2.5.4 scipy==1.18.1 mink==1.3.0 qpsolvers==4.13.0 daqp==0.10.3 rich tqdm loop_rate_limiters imageio
pip install -q --no-deps -e GMR
python3 -c "from importlib.metadata import version as v; print(*[f'{p} {v(p)}' for p in ('torch', 'mink', 'mujoco', 'general_motion_retargeting')], sep=' · ')"Example from the repository
torch 2.14.1+cpu · mink 1.3.0 · mujoco 3.16.0 · general_motion_retargeting 0.2.0
One line with the four versions. GMR installed in editable mode points to the ~/fth/GMR folder.
0410 s of boxing on the G1
retarget.py reads the BVH the way GMR's script for Xsens does (without its graphical interface), calls GMR frame by frame and checks each frame against the limits of the Menagerie G1.
Terminal · retargeting
F=${FTH:-$HOME/fth}; cd $F/corpo-inteiro && . $F/.venv-gmr/bin/activate
python3 kit/retarget.py $F/GMR/assets/xsens_bvh_test/251021_04_boxing_120Hz_cm_3DsMax.bvh --fim 1200 --nome boxe_g1; echo "saída: $?"Excerpt · kit/retarget.py
folga = np.deg2rad(FOLGA_GRAUS)
fora = (q < lo - folga) | (q > hi + folga)Example from the repository
clipe: 251021_04_boxing_120Hz_cm_3DsMax.bvh · 1200 quadros a 120 Hz (10.0 s) · 28 pontos do esqueleto altura do humano no clipe: 1.67 m · robô: unitree_g1 (GMR, config bvh_xsens_to_g1.json) qpos: 1200 × 36 (base 7 + 29 juntas, mesma ordem do Menagerie) menor margem até o limite: right_ankle_roll_joint (0.0°, 0 = encostado) violações de limite (Menagerie, folga 0.5°): 0 tempo de retargeting: menos de 1 min em CPU gravado: ~/fth/movimentos/boxe_g1.npz saída: 0
1200 frames, no violations. The smallest margin is the ankle touching the limit: a flat foot on the ground takes the ankle to the end of its range.
ERRO: o GMR não está neste ambiente (No module named 'general_motion_retargeting'). Ative ~/fth/.venv-gmr (aula 3). saída: 2
You ran it with the MuJoCo .venv active. GMR only exists in .venv-gmr: activate it (. ~/fth/.venv-gmr/bin/activate) and run again.
05Practice
Install, retarget, check.
E2 · Retargeting with no violations
5 minRun the retargeting of the 10 s of boxing.
Done when: python3 kit/retarget.py ~/fth/GMR/assets/xsens_bvh_test/251021_04_boxing_120Hz_cm_3DsMax.bvh --fim 1200 --nome boxe_g1 → violações de limite (Menagerie, folga 0.5°): 0 and saída: 0.
E3 · Another segment
5 minRetarget from frame 1200 to 2400 with --inicio 1200 --fim 2400 --nome boxe_trecho2 and compare the smallest margin with the one from the first segment.
Done when: ls ~/fth/movimentos/boxe_trecho2.npz → the file exists, and the retarget output ends in saída: 0.
06Lesson quiz
The explanation appears after your answer.
Question 1 of 3
Why does the course install GMR with --no-deps?skill
Correct
The code uses the daqp solver; proxqp, smplx and redis are left out. Pinning by hand only what runs makes the environment smaller and reproducible.Not that one
The code uses the daqp solver; proxqp, smplx and redis are left out. Pinning by hand only what runs makes the environment smaller and reproducible.Review: Why a separate environment just for GMRQuestion 2 of 3
The retargeting gave 0 violations against the Menagerie limits. Why isn't that luck?
Correct
On 15 of the 29 joints GMR's model is more restrictive, and the IK respects those limits. In one direction the guarantee holds; in the other, it doesn't. The ankle touches the limit, so the movement is not "far from the limits".Not that one
On 15 of the 29 joints GMR's model is more restrictive, and the IK respects those limits. In one direction the guarantee holds; in the other, it doesn't. The ankle touches the limit, so the movement is not "far from the limits".Review: Why a separate environment just for GMRQuestion 3 of 3
Each frame's qpos has 36 numbers. What are the first 7?
Correct
The floating base has 3 for position + 4 for the quaternion; the kit keeps w, x, y, z as MuJoCo uses (GMR's example script saves x, y, z, w in the pickle).Not that one
The floating base has 3 for position + 4 for the quaternion; the kit keeps w, x, y, z as MuJoCo uses (GMR's example script saves x, y, z, w in the pickle).Review: One retargeting frameQuiz complete
You got {a} of {n} right. If you get them all right, the lesson is marked as complete.Choose an option. The explanation appears after you try.
07In three sentences
Per-frame IKScales the human and solves the robot's pose frame by frame.
EnvironmentOne venv per tool, pinned versions, dependencies set by code.
LimitsCheck it on the model that will run it, not the one that generated it.