CMU projects.

Quality inspection of complex, custom 3D-printed parts, the kind eventually needed for spacecraft and satellites. Coverage planning for a robot arm and turntable so a line-scanner can fully scan an arbitrary object.

Software perception for one of the first truly autonomous micro-rovers targeted for the lunar surface. Worked on the stereo reconstruction pipeline, getting disparity data saved and inspectable so failures could be debugged offline.

A functional three-legged robot dog built to study how a tripod's natural gait compares to a trajectory-optimized gait in cost of transport, on the same hardware, with stability actively managed at every step.

Six projects from CMU's graduate computer vision course: spatial pyramid scene classification, augmented reality with planar homographies, Lucas-Kanade tracking, two-view 3D reconstruction, neural networks for character recognition, and photometric stereo.

Grading diabetic retinopathy from retinal photographs with AlexNet, ResNet50 and VGG16. Detecting the disease turned out to be easy (98%) and grading its severity hard (80%), so we broke the five-way problem into a cascade of binary classifiers to find out why.

Three searches over three different kinds of space: a grid, where the trick is choosing which goal to aim at; a continuous high-DOF configuration space you can only sample; and a space of logical facts where states are sets of conditions.

Lucas-Kanade tracking is a nonlinear least-squares problem solved once per frame, so which optimizer should solve it? Gradient descent, Newton, Gauss-Newton, BFGS and Levenberg-Marquardt raced against each other, explaining why the 1981 paper chose what it did.

Four takes on the same problem: particle filter localization, EKF-SLAM, SLAM as a sparse least-squares problem, and dense SLAM with point-based fusion. From thousands of samples, to a single Gaussian, to batch linear algebra, to raw geometry.

Setting up ORB-SLAM2, RTAB-Map and Google Cartographer in ROS and benchmarking them against the same TUM data and four driving scenarios we generated in the CARLA simulator. Every system degraded in dynamic environments; the interesting part is how differently.