Simultaneous localization and mapping, or SLAM, is a technique robots use to understand their location and pose in an environment. Using inputs from a variety of sensors, which can include anything from GPS to laser rangefinders, a robot will build a map of its environment and try to understand its orientation and location within it. However, to build an accurate map, a robot requires an accurate understanding of its pose, but to accurately understand its pose relative to an environment, it also requires an accurate map. Solving this SLAM problem-optimizing the software and sensors for overall efficiency, accuracy, and cost-has been a continual challenge for robotics researchers.
Current SLAM implementations vary greatly. The DARPA challenge vehicles, for example, use an expensive suite of sensors like GPS, inertial navigation systems, and lidar with powerful processors to navigate their courses. Inexpensive consumer robots like robotic vacuums use commercial cameras or laser rangefinders combined with wheel odometry and low-power computation for basic SLAM. Others systems employ vision systems or sonar, and different statistical methods can be used to arbitrate between sensor inputs. Selecting the correct combination of sensors and statistical methods is a time-consuming process, and researchers often have to reinvent the wheel to develop the approach they believe is best suited to a specific application.
The Rawseeds Project
Roboticists working under the European Union’s (EU) Robotics Advancements through Web-publishing of Sensorial and Elaborated Extensive Data Sets (Rawseeds) project have provided a set of free tools to benchmark SLAM algorithms that they believe will enable researchers to compare the effectiveness of various algorithms in a more efficient and objective manner. This eliminates the need to develop algorithms from scratch for every application. Because the method compares the estimated poses and planned paths of the robot, rather than the map of the environment itself, SLAM algorithms can be compared independently of the sensor suite integrated into the robot. This allows researchers to work toward an optimum solution more quickly.
The benchmarking method is not ideal-it still requires manual entry of data points during benchmarking-so it is likely to remain in the lab for now. However, as the method is refined, it will have a significant impact on rapid and robust development of commercial SLAM systems, and therefore act as a facilitator for advanced automobiles, home robots, and other consumer systems.