---

name: envs-and-transforms description: "Build, validate, transform, vectorize, and debug TorchRL environments and TensorDict transition layouts." disable-model-invocation: true metadata: disco-role: operating license: MIT

TorchRL envs-and-transforms

Load this sub-skill when the task involves TorchRL environments, transformed environments, vectorized environments, environment specs, optional simulator wrappers, or the TensorDict transition layout produced by EnvBase.step and consumed by step_mdp.

Use this skill for

• Building CPU-safe native environments such as PendulumEnv and custom EnvBase subclasses.
• Wrapping Gym/Gymnasium and other simulator backends with TorchRL environment wrappers.
• Adding TransformedEnv, Compose, ObservationNorm, StepCounter, ActionScaling, image transforms, reward transforms, action transforms, and key transforms.
• Checking and debugging observation_spec, action_spec, reward_spec, done_spec, input_spec, and output_spec with check_env_specs.
• Moving a transition from root + "next" layout to the next-step root layout with step_mdp.
• Choosing SerialEnv versus ParallelEnv, debugging vectorized worker issues, and validating specs before multiprocessing.
• Reasoning about multi-agent environment grouping and nested per-agent specs.
• Routing optional simulator, rendering, and pixel dependencies without claiming they are verified in a base CPU environment.

Route away

• Policy, actor, critic, distribution, recurrent module, or network construction: use modules-and-policies.
• Collectors, replay buffers, samplers, writers, or data-collection topology: use collectors-and-replay.
• PPO/SAC/DQN/objective wiring, value estimators, target updates, trainers, Hydra configs, or SOTA recipes: use objectives-and-training.
• VLA-specific action chunk training targets and service-backed robot/LLM workflows: use llm-vla-and-services. Generic ActionChunkTransform key semantics are summarized here only to route safely.

Fast operating path

1. Pick the smallest environment surface: native PendulumEnv or a custom EnvBase for core TorchRL behavior; wrapper classes only when the simulator dependency is installed.
2. Inspect specs before writing policy or collector code. Validate with check_env_specs(env, ...) offline because it runs a short rollout and may reset seeding state.
3. Add transforms with explicit in_keys, out_keys, in_keys_inv, and out_keys_inv; verify transformed specs and use clone() before reusing a transform already attached to another env.
4. Run a tiny rollout, inspect root keys and ("next", ...) keys, then use step_mdp with the env's reward_keys, done_keys, and action_keys when keys are nested or multi-agent.
5. Debug in one process with SerialEnv; move to ParallelEnv only after the same factory passes check_env_specs.
6. For optional Gym/Gymnasium, MuJoCo, DM Control, IsaacLab, VMAS, PettingZoo, OpenSpiel, Brax, Jumanji, rendering, or pixel stacks, follow the optional-backend checklist before treating failures as TorchRL bugs.

References

• Environment workflows: construction, specs, rollout, step_mdp, vectorization, multi-agent layouts.
• Transform reference: TransformedEnv, Compose, key routing, normalization, action scaling, and action chunk routing.
• Optional environment backends: Gym/Gymnasium and simulator extras, rendering, and backend-specific cautions.
• Troubleshooting: common failure signatures and fixes.

Safe smoke helper

Run scripts/smoke_env_rollout.py in an environment where torchrl, torch, and tensordict import. From this sub-skill directory:

python scripts/smoke_env_rollout.py --steps 3 --check-specs

From another working directory, pass the actual local path of this sub-skill's scripts/smoke_env_rollout.py copy to Python.

Expected signal: concise success text showing the TorchRL version, rollout batch size, and step_mdp keys. The helper uses only native PendulumEnv, TransformedEnv, StepCounter, check_env_specs, and step_mdp; it does not require Gym, rendering, multiprocessing, or GPUs.