Train a policy

Before starting: select and check a robot with setup. Configure its Gazebo adapter first if you intend to use Gazebo with this run.

Start with validation enabled

pixi run train --preset navigation

The starter recipe uses 64 environments, a 3,000-iteration budget and validation every 200 iterations. These are starting settings, not a navigation-quality guarantee. Reduce the environment count for a smaller machine:

pixi run train --preset navigation --num-envs 8

The CLI shows model preparation, dataset creation/reuse, training progress and the completed checkpoint. Data and robot assets are frozen into the run.

Inspect or resume

pixi run dwarl runs list
pixi run dwarl runs watch
pixi run train --resume runs/<run> --max-iterations 5000

The resume budget is 5,000 total iterations. Resume uses the saved recipe. Use a new run to change the robot, observations or rewards.

Add --detached for background training or --show for the simulator viewer. Monitoring and resume covers logs, TensorBoard and interruption.

Presets and configuration

smoke is a two-iteration pipeline check; navigation explicitly enables periodic validation. Plain pixi run train retains the normal defaults with intermediate checks disabled. Presets are recorded in the run.

Inspect the resolved settings with --print-config or inputs with --dry-run. Configuration explains YAML, overrides and early stopping.

Next: evaluate a checkpoint. The CLI prints a pinned test command.