Core Concepts

The package is easier to use once you separate the major abstractions:

• Spaces define the structure, dtype, bounds, and backend placement of data.
• Environments define runtime interaction through reset, step, and render.
• Worlds and nodes let you compose an environment out of reusable subsystems.
• Wrappers and transformations adapt an environment or data stream without rewriting the source implementation.
• Batches and replay buffers bring the same typed-space mindset to stored data.

Read These First

• Environments

  Stateful Env, functional FuncEnv, and the adapter between them.

  Open the page

• Spaces and Backends

  The contract that keeps data portable across simulators, wrappers, and storages.

  Open the page

• World Composition

  World, WorldNode, and WorldEnv for building richer environments from parts.

  Open the page

Design Direction

UniEnv keeps two ideas front and center:

• Structure should stay explicit. Observations, actions, contexts, and stored transitions should remain described by spaces instead of becoming anonymous tensors too early.
• Backend choice should stay late. The same logic should be able to target NumPy, PyTorch, JAX, and storage-layer backends with minimal translation code.