rueda

Physical AI Data Platform — unified multi-sensor dataset access, spatially-aware augmentation, and task-specific tensor generation for robotics and autonomous driving.

Installation

pip install rueda

For visualization support:

pip install "rueda[viz]"

Overview

Rueda provides a three-layer pipeline for Physical AI data:

1. WorldDataset — Format-agnostic access to multi-sensor datasets (NuScenes, KITTI, Argoverse2, TUMTraf). Each sample bundles camera images, LiDAR pointclouds, calibration, poses, and 3D annotations into a unified DataSample.

2. Augmenters — Spatially-aware augmentation that respects sensor geometry. Transforms are applied consistently across all modalities using transform-graph for coordinate frame management.

3. Generators — Task-specific tensor generation. Convert augmented DataSample instances into model-ready tensors (BEV grids, voxel occupancy, VLM context).

Architecture

Raw Data → WorldDataset → DataSample → Augmenter → DataSample → Generator → Tensors
                                                                      ↓
                                                              ReasoningDataset
                                                                      ↓
                                                               DataContext → VLM

Reasoning Pipeline

For Vision-Language Models, ReasoningDataset wraps any WorldDataset (or loads lightweight image/video/text datasets directly) to produce DataContext — a structured container with frames, prompts, and ground truth ready for VLM consumption.

Documentation

Full documentation is available at vistralis.org/rueda.

Development

Install in editable mode with development dependencies:

pip install -e ".[dev]"

Linting & Formatting

ruff check .
ruff format .

Testing

pytest
pytest --cov=rueda --cov-report=term-missing

CI/CD

• CI: Lint + test on Python 3.10–3.14 (push to main, PRs)
• Release: Tag v* triggers automated PyPI publish via trusted publishing

License

Apache 2.0 — Vistralis Labs