To find their way around the world, autonomous robots need effective sensing and navigation systems. Some of the best navigation algorithms around rely on the rich environmental data provided by LiDAR-based SLAM (Simultaneous Localization and Mapping) setups. The three-dimensional mapping data provided by these systems give a very clear picture of the world around a robot, which is crucial information when plotting a course.
But all that data comes at a cost. The LiDAR sensors used by autonomous robots send out rapid pulses of laser light and measure the reflections to determine the distance to surrounding objects. Over time, this builds up a dense 3D picture of the world — but processing and storing all that information will eventually consume a large amount of computational resources. Before long, the 3D point clouds collected by the system will hog tens of gigabytes of memory.
In an effort to make the process more computationally efficient, a team led by researchers at Northeastern University developed what they call DFLIOM (Deep Feature Assisted LiDAR Inertial Odometry and Mapping) — an algorithm that dramatically reduces resource usage without compromising accuracy.
Architecture of the feature extraction network (📷: Z. Dong et al.)
DFLIOM builds upon an earlier system called DLIOM, which fuses LiDAR and inertial measurement unit data to estimate a robot’s movement through space. While DLIOM processes either full LiDAR point clouds or uses features selected through manually crafted heuristics (like edges or flat planes), DFLIOM takes a different path. It uses a lightweight neural network to automatically select only the most relevant points from the point cloud, based on their value to SLAM objectives like scan registration and pose estimation.
Rather than relying on simple geometric cues, this deep learning-based approach identifies semantically meaningful features. This may involve ignoring moving objects (like people or cars) or prioritizing static structures (like walls and signs), for instance. The result is a smarter, leaner mapping process.
In tests conducted using an Agile X Scout Mini mobile robot on Northeastern’s campus, DFLIOM reduced memory usage by 57.5% and decreased localization error by 2.4% compared to state-of-the-art methods. It achieved these gains using only about 20% of the original point cloud data — without compromising real-time performance.
By focusing on what matters most, DFLIOM appears to be a promising step toward more efficient, scalable, and intelligent SLAM systems. That could prove to be vital for the next generation of delivery robots, autonomous vehicles, and beyond.
