Staff Deep Learning Engineer, State Estimation (R5785)

Job Description:

What you'll do:

Develop and evaluate models for tasks such as feature detection and matching, visual correspondence, depth estimation, relative pose estimation, and image-to-map localization.
Combine learned visual representations with geometric methods to improve localization accuracy, robustness, and recovery under challenging conditions.
Own data preparation and supervision strategies, including dataset curation, annotation requirements, labeling tools, automated quality checks, and coverage analysis.
Select and integrate deep learning tools and build reproducible training workflows, including experiment tracking, configuration management, and dataset and model versioning.
Design evaluations that measure both model performance and downstream localization outcomes across changes in lighting, viewpoint, altitude, terrain, weather, and sensor characteristics.
Analyze failures and use controlled experiments to prioritize improvements to data, supervision, models, and integration.
Partner with state estimation engineers to integrate learned measurements and confidence estimates into VIO and terrain-relative navigation systems.
Profile models against onboard compute, memory, and latency constraints, and work with deployment engineers on optimization and runtime validation.
Deliver tested, documented components and interfaces for Hivemind SDK, collaborating with software, systems, and flight test teams.

Required qualifications:

3D computer vision and geometric methods: Strong foundations in camera models, coordinate transformations, projective geometry, and multi-view geometry, with practical experience in one or more fields: vision-based navigation, visual geolocation, Structure from Motion (SfM), SLAM, 3D reconstruction, depth estimation or similar fields. Expertise in every area is not required.
Deep learning for computer vision: Hands-on experience designing, training, debugging, and evaluating models using PyTorch or an equivalent framework, including architecture selection, loss design, optimization, and augmentation that preserves geometric consistency.
Mathematical foundations: Solid understanding of linear algebra, probability, and numerical optimization, with the ability to apply these concepts to learning and geometric estimation problems.
Data preparation and supervision: Experience building pipelines for sensor data ingestion, cleaning, filtering, deduplication, and dataset versioning. Ability to define annotation requirements, select or build labeling tools, and assess the quality of annotations, pseudo-labels, and reference measurements.
Training workflows and tooling: Ability to select and integrate development tools and build reproducible training workflows, including configuration management, experiment tracking, checkpointing, and GPU performance troubleshooting.
Evaluation and failure analysis: Experience designing benchmarks, preventing data leakage across related sequences or locations, analyzing performance across operating conditions, and connecting model metrics to downstream geometric or localization accuracy.
Deployment guidance: Ability to profile inference latency and memory use, document model interfaces and preprocessing, assess accuracy–compute tradeoffs, and advise deployment engineers on export, precision, and runtime optimization.
Software engineering and ownership: Strong Python skills and experience writing maintainable, reusable software. Demonstrated ability to take a computer vision capability from problem definition and raw data through training, evaluation, and integration readiness.
Technical collaboration: Ability to communicate assumptions, experimental findings, and design tradeoffs clearly and translate research into working software.
A degree in computer science, robotics, electrical engineering, applied mathematics, or a related field, or equivalent practical experience.

Preferred qualifications:

Visual navigation: Experience with visual odometry, VIO, terrain-relative navigation, map-based localization, or related navigation systems and
Correlation-based methods: Experience with feature correlation, correlation or cost volumes, and matching methods for correspondence, stereo, optical flow, or localization.
Learning strong priors: Experience learning priors over scene geometry, depth, motion, or appearance to improve estimation under sparse, ambiguous, or degraded observations.
Diffusion and flow matching: Experience applying diffusion models or flow matching to computer vision, geometric inference, or conditional generation.
Efficient inference: Experience with model export, quantization, TensorRT, ONNX, or deployment on embedded compute platforms.
You do not need every nice-to-have to succeed in this role. We welcome engineers with strong deep learning and 3D vision foundations who are excited to develop expertise in state estimation and navigation.