Physical AI R&D Engineer

Navflex•Broomfield, CO

31d•$115,000 - $140,000

About The Position

Our objective is autonomy in unstructured, real-world environments: vehicles that build a rich spatial understanding of their surroundings, reason about how they'll change, and act decisively within them. Progress compounds: fleet data becomes the substrate for world modeling and simulation, which lets us validate new behaviors faster than physical testing alone. You'll work on problems where evaluation methodology is still evolving, and you'll contribute to shaping it alongside senior team members. We expect you to track the literature, judge what is ready for real hardware, carry results from prototype to validated capability, and work with core software engineering to bring it into production behind clear requirements and acceptance criteria. Staying current is part of the job, not something you do on your own time. This is research and development, not research alone. We work on open problems, but always against a product roadmap, and a result counts when it reaches a vehicle and changes something a customer experiences. We move quickly, and nobody here disappears into a lab. Focus Area: Spatial AI Perception already runs on our vehicles across LiDAR, depth cameras, inertial sensing, and wheel odometry. You'll deepen its understanding: improve the accuracy and robustness of today's models, harden them against the conditions that degrade performance, and decide what should replace them. The larger goal is a probabilistic 3D representation of the vehicle's surroundings that holds up over time: persisting through occlusion, decaying honestly when the world may have changed since it was last seen, and carrying calibrated uncertainty the rest of the stack can rely on. Warehouses are dynamic: pallets, people, and other vehicles move while unobserved. How it's encoded is an open question you'll help answer. The same work extends into state estimation, where learned components complement the classical stack our autonomy team maintains: visual odometry, degeneracy prediction in geometrically weak environments, and pose uncertainty that's trustworthy rather than optimistic. Success Measures Models you build run on production vehicles, with field behavior backed by evaluation developed in partnership with the team, with your own contributions clearly reflected in the results Capabilities you prototype reach production because you contribute clear, well-documented work that the team can build against You help evaluate external methods and contribute to team discussions on what's worth pursuing

Requirements

MS or higher in Computer Science, Robotics, Computer Engineering, Electrical Engineering, or a related field
You've contributed meaningfully to a research project or thesis, with exposure to presenting and defending results
You stay current with recent literature relevant to your area of focus.
Willingness to seek guidance from senior team members on production-readiness decisions.
Strong Python with modern ML frameworks such as PyTorch, plus working C++ or the clear ability to develop in it
Experience training and evaluating models on real sensor data, not public benchmarks alone, across modalities such as camera, LiDAR, radar, depth, or inertial
Professional fluency in English

Nice To Haves

Depth and current familiarity in one or more of: 3D scene representation: occupancy, scene graphs, or neural and splat-based reconstruction
Segmentation, detection, or other perception on real sensor data, including vision foundation model adaptation
Multi-modal fusion across camera, LiDAR, radar, depth, or inertial data
SLAM, visual odometry, or geometric computer vision
Uncertainty quantification and calibrated probabilistic modeling
Research you took onto physical hardware, from a lab platform to a production system
Experience with ROS and ROS 2
GPU deployment and inference optimization, for example CUDA or TensorRT
Simulation, synthetic data, or digital twins, for example Gazebo, Isaac Sim, Isaac Lab, MuJoCo, or NVIDIA Cosmos, plus sim-to-real practices

Responsibilities

Take promising methods from the literature through prototype to validated on-vehicle capability, then partner with core software engineering to turn a proof of concept into production code
Train, evaluate, and iterate on models using real fleet data: logs, rosbags, and multi-modal sensor streams from vehicles in production
Build and maintain the 3D representation the rest of the stack consumes, with the calibration and sensor health monitoring it depends on
Contribute to baselines and evaluation methodology, applying and refining standards set by the team
Work with core software and test engineering on annotation pipelines, data infrastructure, and simulation tooling