Research

Visual demonstrations of algorithms and field deployments

My research develops algorithms and systems that enable robots to explore and monitor challenging environments, with a particular focus on underwater environments. I work on the tools that will allow future robots to autonomously study and understand the natural world.

For a complete list of my publications, see my Google Scholar profile. Below are some quick highlights from recent projects.

Certifiably Correct Navigation

Two major challenge in robotic deployments stem from modern localization systems. Current localization techniques are not only brittle in the face of real-world conditions, but they often fail silently without any indication and thus no ability to recover. Our work in certifiable navigation aims to address these challenges by developing algorithms that not only are capable of guaranteeing optimal performance under certain conditions, but also provide certificates of correctness that can be used to detect failures and trigger recovery behaviors.

Below you can see our CORA algorithm (a certifiably correct solver) in action on several real-world datasets. We are randomly initializing each instance – a scenario that breaks existing state-of-the-art solvers – and showing that regardless of initialization, our approach returns a certifiably correct estimate of the robot trajectories.

To add to the excitement, we can do all of this faster than the existing methods.

Four different benchmark data sets that our solver (CORA) is solving. Each of these is randomly initialized and the solver progressively "untangles" them and finds the globally optimal solution.
Left: Runtime speedup our method gets over GTSAM, a state-of-the-art solver for robot navigation problems. Right: A fun animation of the spectrahedron, a geometric object closely related to the tools used in certifiably correct optimization.

Large-Scale Multi-Robot Navigation

This is a collection of experiments set up to showcase the benefits of certifiably correct navigation in complex, multi-agent localization problems. We built three sensor payloads, each with a camera (for visual odometry) and ultra-widebad radio (to get distance measurements between the different payloads). The goal was to see if we can use the payloads own (noisy) motion and the inter-agent distance measurements to (a) correct for the error in the odometry and (b) obtain a globally consistent estimate for all of the payload’s trajectories.

Technologies like this are critical for large-scale mapping and observation settings, where many agents will need to be deployed and globally consistent representations will need to be constructed from their distributed observations.

The image stream captured by the three different payloads.
An overlay of trajectory estimates given by visual odometry (VO). Notice the substantial drift given by the VO systems.
Our solver fusing all of the information to obtain a globally consistent and certifiably correct estimate of the three trajectories.

Marine Robotics

Some experiments done in the Charles River in Cambridge, MA on a particularly pleasant day. Featuring an autonomous kayak (the jetyak) and my wonderful collaborator, Brendan O’Neill.