The interface is localized. Skill descriptions and documents are machine-translated by AI into your language; code, commands, paths and technical terms stay in the original, and the author's document governs.
Explore skills
Productivity · SKILL.md

suite-rl-workflows

suite-rl-workflows

Load, inspect, step, wrap, and validate dm_control suite environments.

Independent GitHub snapshot · Automated classification · Original documentation governs

Skill documentation

GitHub

name: suite-rl-workflows description: "Load, inspect, step, wrap, and validate dm_control suite environments." metadata: disco-role: operating disable-model-invocation: true license: Apache 2.0

suite-rl-workflows

Use this skill when the request mentions Control Suite benchmark tasks, suite.load, action_spec / observation_spec, dm_env.TimeStep, random rollouts, flat observations, action scaling or noise wrappers, pixel observations, or reward visualization.

Start here

  • Suite API reference — read this for loader signatures, task collections, TimeStep basics, validation snippets, and the task_kwargs / environment_kwargs split.
  • Wrappers and observations — read this when you need action_scale, action_noise, pixels, mujoco_profiling, flat-observation behavior, or reward-color cues.
  • Troubleshooting — read this when the load, rollout, render, or import path fails.
  • Random rollout script — run this for a safe installed-package smoke test that prints spec summaries and steps a built-in suite task.

What this skill covers

  • dm_control.suite.load(...) and dm_control.suite.build_environment(...)
  • suite.ALL_TASKS, suite.BENCHMARKING, and suite.TASKS_BY_DOMAIN
  • dm_control.rl.control.Environment, Task, Physics, and flatten_observation
  • suite wrappers: action_scale, action_noise, pixels, and mujoco_profiling
  • minimal non-rendering rollout loops and validation signals
  • visualize_reward for rendered reward cues

What to route elsewhere

Quick decision guide

  1. Need a built-in Control Suite task? Use suite.load(domain_name, task_name, ...).
  2. Need to confirm names before loading? Check suite.TASKS_BY_DOMAIN or iterate suite.ALL_TASKS.
  3. Need normalized actions? Use action_scale.
  4. Need exploration noise? Use action_noise.
  5. Need pixels? Use pixels and read the backend caveats first.
  6. Need a flat state vector? Use flat_observation=True or control.flatten_observation.
  7. Need step timing? Use mujoco_profiling.
  8. Need registry-style built-in environments outside the Control Suite benchmark family? Route to locomotion-manipulation.
Saved skills