---

name: training-evaluation-cli description: "Operate mjlab training, playback, export, registry, and motion-data CLIs." metadata: disco-role: operating scenarios: - reinforcement-learning-workflows - embodied-ai-simulation disable-model-invocation: true license: Apache 2.0

training-evaluation-cli

Use this sub-skill when the request is about installed mjlab command-line workflows: task discovery, training, playback, export, debug viewing, motion preprocessing, checkpoint selection, W&B handling, or bounded verification.

Use this sub-skill for

• list-envs and task registry inspection
• train, play, demo, export-scene, and viz-nan
• RSL-RL runner config, checkpoint, and W&B resume handling
• motion imitation CSV validation before conversion
• safe help checks and other bounded smoke tests

Route elsewhere

• Environment config or manager-term wiring: environment-configuration
• Scene/entity/simulation/export API details: scene-simulation-assets
• Action, reward, observation, and command choices: mdp-components
• Sensors, terrain, and domain randomization: perception-terrain-randomization

Bundled references

• Task registry
• Training CLI
• Play, export, and debug
• Motion imitation data
• Troubleshooting

Bundled helpers

• check_task_registry.py
• validate_motion_csv_schema.py

Safe first checks

1. uv run list-envs
2. uv run train <TASK> --help
3. uv run play <TASK> --help
4. uv run export-scene <TARGET> --help
5. uv run viz-nan --help

Treat demo, cloud launchers, W&B downloads, and benchmark jobs as networked or long-running unless the user explicitly asks for them.