A first route into Microduck RL

Start with one training task

The microduck_rl project contains environments for learning movement controllers. Its documented workflow uses MuJoCo-based simulation and PPO, followed by an ONNX export for deployment. Local training requires a CUDA GPU and the uv tool.

Before running a long experiment, read the current README and confirm that your machine matches its prerequisites. A simulator that opens successfully in a browser is not a check that local GPU training is configured.

Find the current environment names

In a checkout with the upstream dependencies configured, the project’s registry command lists available environments:

uv run list-envs

The documented walking example is:

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

This launches GPU work; it is not a command to run on the robot. Check the upstream instructions for suitable settings and hosted options. Hugging Face Jobs is an alternative when you do not have local hardware, and hosted compute may cost money.

Keep an experiment notebook

Record the repository commit, environment name, configuration, seed, training duration and output location. Change one variable between runs. Without that record, an improvement can be difficult to reproduce and a failure difficult to diagnose.

Evaluate more than the most impressive moment in a video. Look at startup, stopping, turning, recovery and repeated attempts. Decide what success means before comparing two runs.

Export through the supported path

Use the project’s export workflow rather than an ad hoc conversion. The upstream exporter includes observation normalization. Policy metadata is described separately in the runtime’s manifest reference.

Preserve the original checkpoint and exported artifact together with your notes. Review compatibility with the current runtime before moving from a successful simulation to hardware. Continue with the policy and deployment references.