Real-Time Robotics Perception Engineer

About the role

Build the Path Forward

At Path Robotics, we’re attacking a trillion dollar opportunity - doing things that have never been done before to support an industry hurting from a lack of skilled labor. Big, hard problems are what Path tackles every day, and our people are our greatest asset to get that job done. Our intelligent, hardworking team of people do the impossible every single day, yet remain incredibly kind, humble, and always ready to support one another.

We are building an elite robotics and perception team, and we are looking for an exceptional hands-on Perception Engineer to own the vision a real-time control loop closes on.

What you'll do

- The real-time perception pipeline, end to end — sensing through extraction and geometry to the output another engineer's control loop acts on, at rate, on embedded hardware, inside a budget you set and measure.

- Robustness where the signal is being actively degraded. Getting a reliable measurement out of imagery that something is deliberately ruining is the hardest part of the job and the part that decides whether the whole system works.

- Making the estimate honest about its own failure. A confident wrong answer is the most dangerous thing this role can produce, because the loop will act on it. We want estimates that know when they are wrong and say so early enough to matter.

- Geometric estimation against a known model — recovering where things are, well enough to work against, from partial and imperfect data, and knowing when the geometry does not actually support the answer you just produced.

- Choosing and proving the instruments. You will evaluate sensing options, run the experiments that settle them, and defend the choice with numbers. This is a real part of the role, not an occasional errand.

- Setting the technical direction as the perception work grows toward learned methods.

Who You Are

- You have closed a real-time loop on what a camera saw, on real hardware, and something physical moved in response. You can state its rate, its measured end-to-end latency, and how you measured it — not the model's inference time, the whole chain.

- First-principles depth in classical vision and 3-D geometry — camera geometry, triangulation, registration and where it fails, calibration, and propagating error properly rather than quoting a single number.

- Real-time segmentation or feature extraction on degraded imagery, deployed rather than benchmarked, and an understanding of why thin and low-contrast structures break the standard metrics and losses.

- Strong, modern, production C++, and ROS 2.

- GPU-accelerated inference on embedded hardware under a hard per-frame budget — and the habit of measuring the tail rather than the mean.

- A physical accuracy claim you can defend — measured against something independent of the system that produced it, with bias reported separately from repeatability.