The Cosmos Engineering team builds the foundational capabilities behind these models. We are hiring a Senior Deep Learning Engineer to own the
data engine and end-to-end training verification behind the 3D spatial reasoning and perception capabilities of these models: 2D and 3D grounding, metric geometry, spatial reference frames, cross-view correspondence, and embodied spatial reasoning. You will decide what the model learns geometry from, prove that it learned it, and work directly with research scientists in Cosmos Lab to turn 3D research hypotheses into measurable capability in shipped models. If you believe frontier model quality is won or lost in the data and the evaluations, this is the seat where that belief does the most work.
What you'll be doing:- Own the 3D data engine for Cosmos spatial reasoning: source, curate, filter, and balance large-scale real-world image and video corpora into vision-language training data with the coverage and diversity that spatial understanding demands.
- Build the annotation and auto-labeling pipelines that produce 3D-grounded supervision at scale, camera-relative 3D boxes, referring and spatial question answering, free space and reachability, ego-, world-, and object-centric reference frames, cross-view correspondence, camera motion, distance and size, and chain-of-thought traces, validated by programmatic and model-based critics.
- Own data quality end to end: semantic deduplication, automated quality scoring for faithfulness, completeness, and correctness, coverage analysis across scene types and reference frames, and the sampling strategies that keep pre-training and supervised fine-tuning mixtures balanced.
- Verify end-to-end model training: run and validate full pre-training and supervised fine-tuning pipelines, guard reproducibility, catch data and checkpoint regressions, diagnose throughput and loss anomalies, and attribute capability changes back to the specific data and recipe decisions that caused them.
- Build and operate the 3D and spatial evaluation suite, public benchmarks such as CV-Bench, BLINK, RefSpatial, VSI-Bench, SPAR-Bench, and RoboSpatial, NVIDIA's VANTAGE-Bench for real-world fixed-camera video understanding, and in-house benchmarks you design with continuous evaluation and full traceability from every reported score back to the exact weights, inputs, configuration, and evaluation code.
- Partner closely with Cosmos Lab research scientists: translate 3D research hypotheses into dataset and ablation experiments, run them at scale, and feed honest results back into recipe and architecture decisions.
- Operate on large multi-node GPU clusters, tuning data throughput, sharding, and dataloader performance so that data is never the bottleneck on a long training run.
- Ship the results into Cosmos releases, open-source datasets and benchmarks where appropriate, and raise the bar for data and evaluation rigor across the team.
What we need to see:- MS or PhD in Computer Science, Electrical/Computer Engineering, Robotics, or a related field, or equivalent experience.
- 12+ years of proven experience building deep learning systems in Python with PyTorch or JAX on Linux.
- Deep expertise in 3D computer vision, multi-view geometry, structure-from-motion or SLAM, depth and camera pose estimation, point cloud processing, or 3D reconstruction with the practical ability to produce and validate 3D ground truth at scale, not just consume it.
- Hands-on experience with vision-language models, including building the training data and evaluations that measurably improve visual grounding and reasoning quality.
- Demonstrated experience building large-scale multimodal data pipelines: distributed video and image processing, deduplication, captioning and annotation, automated quality metrics, and dataset versioning.
- Experience running and validating large model training on multi-GPU, multi-node clusters, with working knowledge of distributed training and sharding strategies such as data, tensor, and pipeline parallelism or FSDP.
- Rigorous evaluation methodology: designing benchmarks that resist gaming, building clean ablations, and reading results honestly enough to kill your own ideas.
- Excellent written and verbal communication, with a track record of partnering effectively with research scientists and translating research direction into engineering execution.
Ways to stand out from the crowd:- PhD and/or publications at CVPR, ICCV, ECCV, NeurIPS, ICLR, or CoRL in 3D vision, multimodal learning, or embodied AI.
- Experience curating web-scale or petabyte-scale video corpora, and building the distributed processing infrastructure behind it using Ray, Spark, Slurm, or similar.
- Familiarity with 3D foundation models and modern auto-labeling techniques for geometry, camera estimation, and correspondence.
- Experience with vision-language model post-training, supervised fine-tuning, chain-of-thought data design, reward modeling, or reinforcement learning with verifiable rewards applied to reasoning quality.
- Experience building evaluation infrastructure and harnesses such as VLMEvalKit, including leaderboards, dashboards, and example-level failure inspection.
Your base salary will be determined based on your location, experience, and the pay of employees in similar positions. The base salary range is 224,000 USD - 356,500 USD.
You will also be eligible for equity and benefits.
Applications for this job will be accepted at least until September 19, 2026.
This posting is for an existing vacancy.
NVIDIA uses AI tools in its recruiting processes.
