IsaacSimApp — Synthetic Data Generation Service for NVIDIA Omniverse Isaac Sim (euROBIN MS3, INRIA-reviewed)

IsaacSimApp is a FastAPI service that drives NVIDIA Omniverse Isaac Sim Replicator headlessly to produce annotated synthetic image datasets for AI training in robotics. It exposes endpoints to load USD scenes and trigger capture with parameters (frames, resolution, camera path), runs a rich library of 20+ image augmentations against the generated frames, and produces dataset archives ready for ingestion by downstream platforms such as the SyntetiQ DaaS portal.

Key components: - dataset_server.py — FastAPI app with /load-stage and /healthz endpoints, strict Pydantic schemas, single-job locking. - load_stage.py — headless USD scene loader and capture executor inside Isaac Sim (also runnable from the CLI). - augmentators/ — pluggable per-frame augmentations (motion blur, glass blur, brightness, sigmoid, Sobel, Canny, depth/normals colorisation, pixellate, CutMix, RGB↔HSV, shot/speckle noise, …). - tools/convert_to_pascal_voc.py and tools/convert_to_yolo.py — bounding-box export converters. - tools/zip_dataset.py — content-hash deterministic ZIP packaging. - gcs_upload integration with optional webhook callback.

Validated quality (carried forward from MS2 baseline): - median BRISQUE = 31.8 (target ≤ 35) — PASS - VOC schema errors = 0 — PASS - detector training with early stopping at epoch 478 — PASS

This archive captures the v1.0.0 release, delivered for Milestone 3 of the euROBIN 3rd Open Call Technology Exchange Programme (Sub-Grant Agreement euROBIN_3OC_8). Hosting Technical Partner: INRIA (Serena Ivaldi). The MS3 deliverable was reviewed and approved by INRIA on 30 April 2026.

The accompanying SyntetiQ stack components are archived on Zenodo under DOIs 10.5281/zenodo.19818727 (DaaS portal) and 10.5281/zenodo.19819084 (RoboLab). The continuously updated source is on GitHub at https://github.com/syntetiq/isaacsimapp.

Released under the Apache License 2.0.

Funded by the European Union — euROBIN, the European Robotics and AI Network (Horizon Europe Grant Agreement 101070596).