Computer Vision Engineer — Perception for Autonomy
Location: [Palo Alto / hybrid]
The role:
We fly drones that inspect real infrastructure. That means reconstructing sites accurately enough to detect change over time, and giving the autonomy stack a picture of the world it can actually act on.
You'll own perception for a moving platform — reconstruction, pose, and the simulated environments we use to train and evaluate flight behavior. You'll work closely with the autonomy side without owning the flight controller.
What you'll work on:
- Reconstruction — Gaussian splatting and photogrammetric pipelines producing metrically accurate, georeferenced scenes from drone imagery
- Pose and state estimation — bundle adjustment, RTK/GNSS and IMU fusion, visual-inertial odometry, multi-camera calibration
- Simulation for autonomy — turning reconstructions into training and evaluation environments for flight policies, and characterizing where sim diverges from reality
- Change detection across reconstructions separated by weeks or months
- Perception in the loop — defining what reconstruction and detection deliver to planning, and what happens when the estimate degrades
- Detection and auto-labeling models running on the aircraft under real latency and power budgets
What we need:
- 2+ years in computer vision or robotics perception, with systems that ran outside a lab
- Solid multi-view geometry — you can reason about what your estimator is doing and debug a bundle adjustment that won't converge
- Hands-on SLAM, SfM, or visual-inertial odometry
- Strong PyTorch; real experience training and debugging models on field data that doesn't look like the benchmark
- Have worked on a moving platform — drone, vehicle, or robot — where ground truth is expensive and failures happen on site
- Comfortable at the hardware boundary: camera sync, calibration rigs, reading flight logs
- Enough robotics literacy to talk to the autonomy team — you know what a planner needs from perception and why latency and failure modes matter to it
- Writes clearly enough that another team can act on your design doc
Strong signals:
- 3DGS or NeRF, especially large outdoor scenes
- Reconstruction-backed simulation for robot training
- Sim-to-real transfer or learned dynamics
- ROS/ROS2, PX4/ArduPilot exposure
- C++ alongside Python
- Thermal, depth, or lidar fusion
How we work:
Small team, high autonomy, short path from prototype to field trial. Direct access to real aircraft and real customer sites. We hire people who go find the failure themselves.
