• Michael Kaess
• Publications
• Projects
• News
• Software

Projects

Here are some of my current and past projects at MIT. For my earlier work, please see my Georgia Tech web page.

Kintinuous: Live Dense Mapping

Kintinuous provides live dense 3D mapping of extended environments using RGB-D sensors such as Microsoft's Kinect. Unlike the original KinectFusion algorithm, we are not restricted to a small volume by the available memory on the GPU. Instead, the volume moves through space along with the observer, and a triangular mesh is created for slices that leave the volume. We also combine ICP-based tracking with dense color information to achieve robust tracking in a large variety of environments. Kintinuous has mostly been implemented by Tom Whelan, a gifted GPU programmer and PhD student at the National University of Ireland Maynooth under the supervision of John McDonald. An initial publication has appeared in the RGB-D workshop at RSS 2012.

Building-Scale Lifelong Mapping

In recent work we created a map of our ten floor building, the Stata Center. We used stereo vision and odometry from our PR2 robot to create the map. Keeping the optimization efficient during long-term operation is a challenging problem that we approach using the reduced pose graph formulation by Hordur Johannsson. His animation below shows the map being created for a significant portion of the building, covering a cumulative 10 hours of operation during which the robot traveled 11 km through the building. An initial publication has appeared in the long-term operation workshop at RSS 2012.

iSAM2 and the Bayes Tree

Exploring the connection of sparse linear algebra and graphical models yielded new insights into the probabilistic interpretation of matrix factorization, formalized in the Bayes tree (WAFR10 paper). Based on this insight, we developed a novel incremental nonlinear optimization algorithm for sparse least-squares problems called iSAM2 (IJRR12 paper). Our newest development is an incremental version of Powell's dog leg method that provides guaranteed convergence (at least to a local minimum) and extends iSAM2 to non-Gaussian noise models (ICRA12 paper).

Ship Hull Inspection

We integrated localization and mapping with closed loop control of an autonomous underwater vehicle for in-water inspection of large ships. In addition to standard sensors of mobile robotics such as a ring laser gyro, our Hovering Autonomous Underwater Vehicle (HAUV) also includes a doppler velocity log (DVL) and an imaging sonar. This work is funded by the Office of Naval Research (ONR). This is joint work with Franz Hover in Ocean Engineering and Ryan Eustice at the University of Michigan, Ann Arbor, and involves several students: Hordur Johannsson, Brendan Englot and Ayoung Kim. Our IJRR journal submission just got accepted, more details soon. Further links can be found here.

Visual SLAM

Multi-session visual mapping in large-scale environments using anchor nodes to connect multiple pose graphs (ECMR11 paper). One motivation is robustness to failures by restarting and eventually merging with the original map. A key component of the system is an improved appearance based loop closure technique. This is joint work with John McDonald at the National University of Ireland, Maynooth, and Jose Neira and Cesar Cadena at the University of Zaragoza, Spain.

All Source Positioning and Navigation (ASPN)

This is a joint project with Georgia Tech and Sarnoff/SRI International sponsored by DARPA. Its goal is to combine low latency filtering methods with the loop closing capabilities of current smoothing solutions to the SLAM problem. The Bayes tree has proven essential for the algorithmic development, and iSAM2 is also being used. Here are some initial publications on concurrent smoothing and filtering and on smoothing with IMUs.

Cooperative Mapping and Localization

We introduced the concept of anchor nodes for combining initially independent pose graphs of multiple robots upon pair-wise encounters. One application was the cooperative mapping of a large underground loading dock by a quadrotor and a ground robot. A paper with my student Been Kim appeared in ICRA 2010. The concept of anchor nodes has since found a range of applications, including for visual SLAM (see above).