Adam Srebrnjak Yang

November 13th, 2026 – 12:00-1:00 pm, EC4-2101A

Monocular SLAM suits small drones, but drones fly exactly the motions that break it. This talk covers the data capture and evaluation carried out as part of developing SLAM initialization methods.

We built ANTARES, a 1.5 kg quadrotor that records the same flight through three cameras and the PX4 IMU at once, and we discuss the priorities and challenges we faced in developing its data recording system.

We then present motion profiling of the flights and the common visual failure modes they expose in visual SLAM datasets. To validate these results, we compare the performance of SLAM systems on our dataset with their performance on the KAIST VIO dataset.

We also collected data with a DJI Mini 5 Pro on two flights of 2.13 km and

6.65 km, and examine how well monocular SLAM systems survive over these distances. We compare the direct, indirect, hybrid, and learned classes of systems on this data.

Finally, we present the failure modes that occurred during these flights and the mechanisms by which each class of SLAM system fails.