Correction, October 20, 2026: An earlier version described this hardware stack as a 2018 configuration. The autonomy project began in 2018, but Intel announced the T265 in January 2019 and NVIDIA announced the Jetson Nano in March 2019. The stack below was a later project iteration.

That iteration used an Intel T265 for visual-inertial odometry, a D435 for depth, an NVIDIA Jetson Nano running ROS, and a PX4 flight controller. An EKF fused the T265 estimate with GPS.

This is a retrospective engineering note; the original flight logs are not public. The speed and timing figures come from our project notes.

The system could localize, build an obstacle representation, plan, and fly. Its safe forward velocity in our tests was still only about 0.5 m/s.

The symptom looked like state estimation

As speed increased, corrections to the fused position invalidated more of the work already done by mapping and planning. The visible symptom was a state estimate that moved, so we spent time retuning the filter, changing noise assumptions, and trying alternative process models.

Those changes affected the estimate, but they did not make 1 m/s a safe operating point for this stack.

The limit was end-to-end

The relevant timing budget covered the complete path from sensing to control:

The T265 estimate drifted between GPS corrections.
GPS arrived at 5 Hz.
The Jetson Nano ran depth processing, occupancy mapping, planning, and the control loop on the same compute budget.

Each stage was usable when measured on its own. In the complete system, a localization correction could force the obstacle map or trajectory to be recomputed. At higher speed, the drone travelled farther while that happened, reducing the distance available to perceive an obstacle and respond to it.

The useful metric was the age of the information by the time a control action reached the aircraft, not the runtime of any one node.

The hardware trade-off

Moving to a more capable Jetson platform would have created more room for perception and planning. It also changed the cost, weight, and power budget enough to affect the aircraft design.

The 0.5 m/s ceiling was therefore not evidence that the EKF itself had failed. It was the operating point at which that sensor, compute, and planning stack still had enough end-to-end margin in our tests.