* Visual SLAM and 3D reconstruction of street scenes using onboard cameras* Multi-lidar SLAM and multi-sensor calibration* Tech lead for onboard calibration of 10 cameras on the Gen2 R1 platform.

Computer Vision, State Estimation, Visual-Inertial Odometry, SymforceWorked on improving the initialization and high-altitude operation of the VIO system. One of the regular maintainers of the Symforce framework.https://github.com/symforce-org/symforce

Computer vision, state estimation, and deep learning in robotics. Advised by Prof. Camillo J. Taylor.* Low-latency, memory-efficient lidar odometry. Fully parallelized lidar odometry with less than 10Mb of fixed onboard memory usage. No dynamic allocation at runtime. Able to achieve sub-millisecond runtime per frame (64x1024 point cloud). Wrote an ouster lidar ros driver to support streaming output, 10x faster than the official ouster driver.https://github.com/versatran01/llolhttps://github.com/versatran01/rofl-betahttps://github.com/KumarRobotics/ouster_decoder* Depth completion via Deep Basis Fitting with uncertainty. A sparsity-aware depth completion module that can be added to almost any depth completion network. A differentiable (non-linear) least squares fitting module is used to estimate the coefficient of depth basis to reconstruct the observed depth map. The bayesian extension can learn a prior on the weights and therefore handle very sparse or even no depth measurements.* Direct visual odometry. A fast and memory-efficient version of DSO, with support for monocular, stereo, and RGB-D cameras. Similar to LLOL, the system is also fully parallelized with less than 10Mb of fixed onboard memory usage. No dynamic allocation at runtime. Able to achieve sub-millisecond frame tracking and <5ms sparse direct bundle adjustment.

CIS240 Introduction to Computer SystemsCIS390 Robotics: Planning and PerceptionMEAM620 Advanced Robotics

Application of computer vision and robotics in precision agriculture.