Dataset Integrations

UniEnv can wrap external dataset ecosystems behind the same BatchBase interface used by replay buffers and storages.

That means downstream code can often stay unchanged whether its data comes from:

• a live replay buffer
• a local dataset on disk
• a Hugging Face dataset
• a PyTorch Dataset

Hugging Face Datasets

HFAsUniEnvDataset adapts a datasets.Dataset into a UniEnv batch.

Highlights:

• infers a UniEnv space from the first sample
• configures the Hugging Face dataset format to match the requested backend
• supports column selection through the batch column-view helpers

Use it when your training or evaluation data is already managed through the Hugging Face datasets stack.

LeRobot Datasets

LeRobotAsUniEnvDataset reads local or Hub-hosted LeRobot datasets without requiring the upstream lerobot Python package.

Highlights:

• supports multiple LeRobot dataset versions
• reconstructs a DictSpace from dataset features
• can include or exclude named features
• can optionally decode video features

This is the most direct path for bringing LeRobot-formatted robotics datasets into UniEnv-native tooling.

PyTorch Dataset Adapters

UniEnv provides both directions:

• PyTorchAsUniEnvDataset: wrap an existing PyTorch Dataset as a UniEnv batch
• UniEnvAsPyTorchDataset: expose a UniEnv batch as a PyTorch Dataset

This is useful when the training stack expects PyTorch dataloaders but the storage or replay layer is written against UniEnv.

Choosing An Adapter

Use a UniEnv adapter when you want:

• one indexing interface across multiple dataset sources
• space-aware structured samples instead of ad hoc dictionaries
• the ability to plug the result into UniEnv samplers or downstream wrappers

Use the original upstream dataset API directly when you need ecosystem-specific features that are outside the UniEnv batch contract.