Vishnu Desaraju

This was my PhD thesis work at CMU. The underlying idea is to leverage past experiences to reduce the computation needed to generate new control commands in a Nonlinear Model Predictive Control (NMPC) formulation. We also consider online dynamics model learning for adaptation to unmodeled disturbances and robustness to imperfect state feedback. This enables the application of high-rate, adaptive NMPC to severely compute-constrained platforms.

See the thesis and our ICRA 2016, ICRA 2017, and RSS 2017 papers for more details. This work was also featured on TechCrunch.

The objective of this program (in collabration with JPL) was to enable a size-weight-and-power (SWaP) constrained aerial robot to autonomously land on a rooftop to conserve power. The integrated perception-planning system runs onboard and identifies viable landing sites on an unknown rooftop while intelligently generating trajectories that automatically balance exploration for site detection and focused landing to conserve power.

See our RSS 2014 and AURO 2015 papers for more details.

As a Guidance, Navigation, and Control engineer at Aurora Flight Sciences, I primarily worked on programs related to coordinated search, track, and patrol using multiple autonomous systems (air, ground, and/or water). This included frequent field tests to evaluate system performance in scenarios with more realistic and challenging conditions (e.g., flying in wind, vision based tracking with varying illumination, communication over large distances, accounting for GPS accuracy).

The Aurora website had a nice overview of several of these programs: OPS-USERS, UAV/USV, and Sensor Fusion (see this web archive link).

This project started as a combination of my machine learning course final project and a labmate's tech report. We applied ideas from machine learning to anticipate driver behavior in order to alert other drivers if a car is expected to violate intersection rules (e.g., run a red light). We trained and tested the classifiers using a real-world dataset with thousands of intersection approaches and were able to predict violations with high accuracy. This work was featured on the MIT homepage as well as dozens of major news sites. It also led to a patent.

For technical details, see our ITS 2012 paper.

This was my Master's thesis work at MIT. The Decentralized Multi-Agent RRT algorithm enables a team of robots to plan conflict-free paths online. Agents compute new plans asynchronously while bidding for a virtual token. Each agent bids according to its potential improvement in path cost, and only the winner may update its plan. Agents coordinate by broadcasting a minimal representation of updated plans. An extension to DMA-RRT introduces pairwise cooperation to avoid common deadlock scenarios.

For details, see the thesis or AURO 2012 paper.

The Agile Robotics project was a spinoff of MIT's DARPA Urban Challenge entry and focused on developing an autonomous forklift that could work in minimally-prepared environments (such as a supply depot on a forward operating base). I led the development of a small support robot (Rover) that had the same autonomous navigation capabilities as the forklift but was safe to take on a guided tour of a supply depot. The idea was that these small robots would be able to traverse the depot more regularly than the forklift and relay updated maps and inventory data.

For more information on Rover, see Chapter 5 of my Master's thesis.

► Video 1: Autonomous navigation ► Video 2: Guided tour via person-following

FIDO was a labmate's Master's thesis project that I spent most of my first summer at MIT helping with. The goal was to have the car autonomously follow and catch a bouncing ball on a specific bounce, without stopping. Both the ball and the car were tracked via motion capture, although the ball required manual data-association. The funnel on the car (actually the top half of a Gatorade bottle) was only twice the diameter of the golf ball, requiring very accurate control with a vehicle that would skid.

For more information, see Sergio Cafarelli's thesis.

My undergraduate research work at Michigan focused on collision avoidance for autonomous vehicles as they merge into a single lane. The general idea developed by the graduate student was to identify states where the vehicles would inevitably collide (a 'capture set') and apply specific control actions that ride the boundary of the capture set to avoid collision. I primarily led the implementation and experimental validation of the approach, and generated the first set of experimental results to come out of this lab.

See our ICRA 2009 paper for details.