Quickstart

Quickstart#

Train and watch your first whole-body tracking policy on the bundled sample clips. Every step works in the Studio (browser) or on the command line — pick a tab once and the others follow.

Before you start

  • wbc-mjlab installed (Installation) — uv run wbc-mjlab-list-envs lists Wbc-G1 and friends

  • An NVIDIA GPU for training (CPU is enough for steps 1–2)

  • About 15 minutes for a smoke run; a real policy trains for hours

No GPU? Try the live demo or Colab instead.

1. Open the pipeline#

uv run wbc-mjlab-studio

Open the printed link. The top bar shows the pipeline: Robot → Data → Preset → Train → Play. G1 is selected.

Everything below runs from the repository root.

2. Look at the sample clips#

The 10 bundled G1 clips (walk, run, dance, fall and get up, flips …) are CSV files. Training reads them as they are — resampling, velocities and body poses are computed on the GPU while it runs — so there is nothing to convert.

Data → pick samples → click a clip to watch it; the Clip tab lists anything suspicious (feet in the floor, joint limits).

Nothing to do: data/g1/samples/ holds the CSVs training reads.

3. Train#

Wbc-G1 is the default stack (which task?). A smoke run of 1000 iterations shows the loop works; a full run is about 30k.

Preset → Wbc-G1 → Train → set iterations to 1k → Launch. The job shows progress, curves and a link to TensorBoard.

uv run wbc-mjlab-train --task Wbc-G1 --dataset samples --agent.max-iterations 1000

Check: logs/rsl_rl/wbc_g1/<run>/ fills with model_*.pt checkpoints and params/policy.onnx + params/config.yaml.

4. Watch the policy#

Play → pick the run → Load policy. Each env shows its reference as a ghost; pick a clip in the bar above the 3D view, push the robot with p.

uv run wbc-mjlab-play --task Wbc-G1 --dataset samples --viewer viser

Open the viser link; the sidebar shows the clip, reference alignment and the adaptive sampling bins.

After 1000 iterations the robot follows the easy clips roughly; flips and get-ups need the full run.

What you have now#

logs/rsl_rl/wbc_g1/<run>/
  model_*.pt              checkpoints (continue with --agent.resume True)
  params/policy.onnx      the policy, ready for a robot runtime
  params/config.yaml      observation layout, joints, PD gains for that runtime

Next steps#

Your own clips

Import and check a motion library.

Prepare motion data
A full training run

Iterations, resume, many GPUs, logs.

Train a policy
Other tasks

ZEST, state estimation, end-effector command, rough terrain.

Choose a task
On the robot

Hand policy.onnx + config.yaml to a deploy runtime.

Export and deploy