Voxel51
Visual-data curation and physical AI · Infrastructure
How they use Nebius AI Cloud
Voxel51 delivers a synthetic-data pipeline for Porsche Research with NVIDIA and Nebius. Teams start in FiftyOne by finding gaps in driving-scene data. Porsche’s AV team uses embedding search to identify scenes with the greatest impact on model performance. They generate scene variations on Nebius infrastructure, score the outputs and bring accepted material back into FiftyOne for review. Nebius describes this as a production pipeline. Nebius reports that the automotive team generated and quality-checked hundreds of realistic scene variations, reducing the time needed to create model-ready training data.
The pipeline covers scene labelling, prompt variation, video generation and quality scoring. It uses NVIDIA Cosmos Reason or Qwen3-VL for auto-labeling and Cosmos Grader for output scoring. The Nebius-backed generation pipeline supports augmentations across weather, lighting and traffic states. Voxel51 describes this as a reference architecture connecting real and synthetic data, supplying the data workflow and evaluation layer while Nebius supplies compute. The implementation combines NVIDIA OSMO orchestration with Cosmos models. Nebius-backed data agents generate prompts from user-defined weather distributions to augment thousands of videos.
NVIDIA names Voxel51 as an early user of its Physical AI Data Factory Blueprint on Nebius infrastructure. CEO Brian Moore describes Nebius supplying compute for users’ auto-labeling and novel-scene generation. FiftyOne workflows on Nebius GPU clusters support dataset curation, augmentation and quality checks. Joint workshop agendas also specify Nebius Serverless AI Jobs for embedding computation and YOLO11n fine-tuning, including the planned October 28, 2026 San Francisco session. The hands-on workflow uses DROID robotics recordings in FiftyOne’s native multimodal MCAP viewer, with detector predictions reviewed on the original recording timeline. The planned Montreal workshop uses embeddings of curated DROID frames for similarity search and visualization. Voxel51 also reports that developers and ML teams have used Nebius AI Cloud at its joint events for multimodal workflows including embeddings and auto-labeling. The workshop repository specifies a Nebius L40S GPU for presenter-led YOLO11n training. Attendees can follow locally on CPU using precomputed GPU results.
Sources
11 sourcesClosing the physical AI learning loop: How Voxel51 is making data quality the competitive edge
19 Oct 2026 · nebius.com · Web page
Nebius describes a production synthetic-data pipeline delivered with Voxel51 and NVIDIA for an automotive manufacturer. The team generated and quality-checked hundreds of realistic scene variations using FiftyOne and Nebius AI Cloud, reducing the time needed to create model-ready training data.
Open original sourceGeekWire Studios: Nebius at NVIDIA GTC 2026 | Dan Gural, Voxel51
9 Apr 2026 · www.youtube.com · Video
Voxel51’s Dan Gural describes a synthetic-data pipeline combining NVIDIA OSMO, Cosmos models and Nebius compute. Nebius-backed data agents generate prompts from user-defined weather distributions to augment thousands of videos. He also outlines future ambitions for digital twins, simulations and VLA training. Evidence at 02:32, 02:47, 05:39, 06:56.
Open original source at 02:32Nebius physical-AI announcement
16 Mar 2026 · Nebius · Web page
Nebius describes Voxel51 customers running FiftyOne workflows on Nebius GPU clusters for dataset curation, augmentation and quality checks. Brian Moore says Nebius supplies compute for auto-labeling and novel-scene generation. The announcement also describes their synthetic-data pipeline for Porsche.
Open original sourcePhysical AI workshop: 10 November 2026 in Montreal
17 Sept 2026 · Voxel51 · Web page
Voxel51 and Nebius co-present the planned November 10, 2026 Montreal workshop. Its agenda uses Nebius Serverless AI Jobs to compute embeddings of curated DROID robotics frames for similarity search and visualization and to fine-tune YOLO11n. Predictions are reviewed on the original MCAP timeline.
Open original sourceGitHub - danielgural/physical-ai-workshop: Nebius x Voxel51 Physical AI Workshop: DROID robot recordings -> FiftyOne -> YOLO11n on Nebius · GitHub
19 Oct 2026 · github.com · Web page
Daniel Gural’s companion repository for the Nebius and Voxel51 workshop specifies Nebius Serverless AI Jobs for presenter-run GPU steps and a Nebius L40S GPU for YOLO11n training. Attendees follow locally on CPU with precomputed GPU results.
Open original sourceCompany context
voxel51.comVoxel51 makes software for preparing visual and multimodal data used to train and evaluate AI models. Its FiftyOne open-source toolkit and enterprise platform help machine-learning engineers and data scientists search, visualize, curate and annotate images, video and sensor recordings, and investigate model failures. Customers work in robotics, autonomous vehicles, manufacturing, healthcare and other fields using computer vision. Its Physical AI Workbench validates and enriches sensor data and connects it to NVIDIA reconstruction and simulation tools, while Voxel51 Agent automates dataset tasks.
Recent updates
7 updatesIntroducing Voxel51 3.0: The end-to-end multimodal platform for physical AI
Voxel51 3.0 unifies search, curation, annotation, and evaluation of multimodal data for physical AI, with Voxel51 Agent and free self-service access.
voxel51.com announcementVoxel51 2.25 adds segment search and a built-in data agent
Voxel51 2.25 introduced embeddings over time windows in sensor recordings, letting users find similar moments and jump directly to them. The release also integrated FiftyOne Agent into the enterprise app and added annotation metrics and custom workflow stages.
Voxel51Voxel51 adds LeRobot dataset support
Voxel51 added LeRobot support to its commercial platform for robotics data.
Voxel51 announcement