Velocity — walking on velocity commands

Mjlab-Velocity-Flat-MicroDuck
Mjlab-Velocity-Rough-MicroDuck

The main task. Start here.

It trains walking to velocity commands while also accepting head-pose commands — what the left and right sticks do in the pad’s default mode.

Training it

uv run train Mjlab-Velocity-Flat-MicroDuck --env.scene.num-envs 4096

The documented reference is 1–2 hours for a usable gait at 4096 parallel environments.

Flat is level ground, Rough is uneven. Train Flat first — if the gait is unstable on level ground, rough terrain only makes the cause harder to locate.

Why it is the main task

Every other task is a skill for a specific situation. This one is everyday driving, and the walking policy the robot ships with comes from here.

It is also the way into the observation contract: all policies share the same 61 dimensions, and this task uses the velocity and head-command parts of it fully.

Deployment

uv run scripts/export.py Mjlab-Velocity-Flat-MicroDuck --wandb-run-path <...>

Point [policy] walk in robotd.toml at the exported .onnx.