Robust and Efficient Robotic Mapping

Cars that drive themselves, submersibles that track the effects of climate change on coral reefs, robots that deliver supplies, aerial vehicles that monitor crops: these are just a few applications that rely on robotic mapping. In some cases, producing a map is the primary goal; in others, the map is needed to support other capabilities such as obstacle avoidance and path planning.

Despite the central role of mapping, robustly and efficiently computing maps remains an open problem. Tremendous progress has been made, yet existing methods struggle to produce good maps when the sensor data is very noisy or when the environment is large. Moreover, despite expending significant computational costs, current methods can produce incorrect maps. Many methods are vulnerable to the non-linearities in map optimization and can get stuck in local minima. Using inexpensive cameras such as cameras exacerbates the problem by increasing noise.

We present a map optimization algorithm that is dramatically faster and more robust to noise than existing methods. Our method is based on stochastic gradient descent, which rapidly explores the search space and can thus escape local minima. We also illustrate the inadequacy of chi-squared error as a performance metric and propose a complementary metric that is often more representative of map quality.

We will demonstrate failure modes for existing methods including the Extended Kalman Filter, Gauss-Seidel relaxation, and Square-Root SAM, and show that our method is significantly more robust.

In addition to synthetic datasets, we present results on benchmark laser-scanner datasets and a 1.9km vision-based dataset that was collected by MIT's DARPA Urban Challenge vehicle.
Edwin Olson is a Ph.D. candidate in the Computer Science and Artifical Intelligence Laboratory at the Massachusetts Institute of Technology. He also received his MEng and BS from MIT. His research spans much of robotics, including machine perception, clustering, robust outlier rejection, planning, and map building. He was a student technical lead on MIT's DARPA Urban Challenge. He is an award winning teacher and the developer of a hands-on autonomous robotics class for undergraduates.