Synthetic data tools for vision & robotics
Updated 2026-10-12
Programmable generation libraries
A programmable Omniverse extension for teams building their own synthetic-data generation and annotation pipeline. It is a component used in NVIDIA workflows, including Isaac Sim, rather than a separate robot simulator.
- Replicator is documented as a core Omniverse extension for building custom synthetic data generation tools and pipelines.
Robot simulation & sensor outputs
A robotics simulation framework for assembling robots, sensors and environments and generating perception data. Its synthetic-data workflow uses the related Replicator tooling; check both when planning annotations and exports.
- Isaac Sim documents RGB, bounding-box, instance-segmentation and semantic-segmentation annotations, with COCO and KITTI export formats.
Procedural generators — code projects
A procedural Blender pipeline for scripting scenes, cameras and materials and rendering labeled training images. Useful when your team wants code-level control over generation and output.
- BlenderProc documents RGB, stereo, depth, normal and segmentation rendering, and writers for HDF5, COCO and BOP annotations.
A research code pipeline for multi-object videos with physics and rich annotations. It supports dataset experiments rather than a hosted Google product, and its setup instructions should be checked before adoption.
- Kubric uses PyBullet for physics and Blender for rendering and produces multi-object videos with segmentation, depth and optical-flow annotations.
A procedural world and asset generation project with documented ground-truth and export workflows. Start here when dataset variety depends on generating geometry and environments in code.
- The Infinigen repository links instructions for camera configuration, extended ground truth, asset generation and exports to formats including OBJ and OpenUSD.
Commercial sensor-data platforms
A commercial synthetic simulation provider with in-cabin and other perception applications. Review its sensor-specific documentation and request output details for your own dataset requirements.
- Anyverse describes its platform as combining a physics-based simulation engine with procedural data generation for perception-system applications.
A reconstruction and sensor-simulation provider for teams starting from captured drive or flight logs. Its documented workflow creates reusable scenes and controlled sensor observations rather than an off-the-shelf image collection.
- Parallel Domain describes camera, lidar and radar simulation, reconstruction from capture logs, and controlled changes to lighting and agent trajectories.
A synthetic-data platform for controllable scenarios and labeled sensor outputs. Stardust is Bifrost’s data product; its Manifold robot-policy evaluation product serves a different task.
- Stardust lists RGB, infrared, depth and segmentation modalities; its product page separately labels radar as coming soon.
Dataset management & augmentation
A platform and service provider for configuring synthetic-data generation workflows, managing assets and iterating datasets. Check the linked product information for simulator integrations and deployment choices.
- Rendered.ai describes an open platform framework with configurable generation workflows, asset management, model training and validation tools.
A data augmentation service for expanding recorded driving imagery into controlled variations. This entry concerns the perception-data workflow; it does not imply vehicle certification or safety approval.
- DriveMatriX documents variations of recorded driving videos across weather and lighting, with changes to lane markings, pedestrians, signs and vehicles.