UniEnv

UniEnv is a Python framework for building robot environments and robot data pipelines on top of a shared set of abstractions.

Documentation: https://unienvorg.github.io/UniEnv/

It gives you:

• a common environment interface for simulation and real-robot control
• a functional environment variant for explicit state passing
• backend-aware spaces, wrappers, and transformations
• world and node composition utilities for multi-component environments
• replay buffers, storages, samplers, and dataset adapters for offline data workflows

The project is designed around one idea: environment code and dataset code should not have to be rewritten every time the simulator, robot, or tensor library changes.

What Is In The Package

• unienv_interface: environments, worlds, nodes, spaces, wrappers, and transformations
• unienv_data: batches, replay buffers, storage backends, samplers, and dataset integrations
• backend portability through XBArray and DLPack-based conversion paths

Installation

Install the base package:

pip install unienv

Install optional extras when needed:

pip install "unienv[gymnasium,video]"

Some integrations and storage backends rely on their own ecosystem packages such as pyarrow, h5py, datasets, huggingface_hub, torch, or jax. Install the ones that match the features you plan to use.

Quick Example

The example below shows the core building blocks without depending on a specific simulator:

import numpy as np

from unienv_interface.backends.numpy import NumpyComputeBackend
from unienv_interface.space.spaces import BoxSpace, DictSpace
from unienv_interface.transformations import RescaleTransformation
from unienv_data.replay_buffer import ReplayBuffer
from unienv_data.storages.parquet import ParquetStorage

backend = NumpyComputeBackend

transition_space = DictSpace(
    backend,
    {
        "obs": BoxSpace(backend, 0.0, 1.0, np.float32, shape=(4,)),
        "action": BoxSpace(backend, -1.0, 1.0, np.float32, shape=(2,)),
        "reward": BoxSpace(backend, -np.inf, np.inf, np.float32, shape=()),
    },
)

action_transform = RescaleTransformation(new_low=0.0, new_high=1.0)
normalized_action_space = action_transform.get_target_space_from_source(
    transition_space["action"]
)

buffer = ReplayBuffer.create(
    ParquetStorage,
    transition_space,
    cache_path="cache/demo_buffer",
    capacity=10_000,
)

print(normalized_action_space)
print(buffer.single_space)

For online logging, append one sample at a time with ReplayBuffer.append(value) and close each segment with mark_segment_end(). Segment visibility depends on the storage backend: legacy storages may expose writes immediately, while VideoStorage finalizes its active segment before it becomes readable.

In practice, you would pair these components with your own Env, FuncEnv, World, or WorldNode implementations, then add wrappers and storage backends as needed.

Documentation

Documentation: https://unienvorg.github.io/UniEnv/

Start with:

• docs/getting-started.md for installation and the package map
• docs/concepts/ for the core abstractions
• docs/guides/ for wrappers, replay buffers, and dataset integrations

Development

For local development:

git clone https://github.com/UniEnvOrg/UniEnv
cd UniEnv/UniEnvPy
pip install -e .[dev,gymnasium,video]
pytest

If you want to exercise optional backends or integrations during development, install their matching dependencies explicitly before running the relevant tests.

Cite

If you use UniEnv in research, cite:

@software{cao_unienv,
  author = {Yunhao Cao AND Cory Fan AND Meryl Zhang AND Sabrina Liu AND Kuan Fang},
  title = {{UniEnv: Unifying Robot Environments and Data APIs}},
  year = {2025},
  month = oct,
  url = {https://github.com/UniEnvOrg/UniEnv},
  license = {MIT}
}

Acknowledgements

UniEnv is influenced by Gymnasium and OpenAI Gym, and builds on ideas from the Array API Standard and DLPack.