LoCal-RIO: Radar-Inertial Odometry with Loop-Closure IMU Bias Calibration

Autonomous robots, from delivery bots to self-driving cars, rely on a constant, internal sense of where they are and how they are moving. In clear, sunny conditions, cameras and laser scanners excel at this task, painting a detailed picture of the world to guide the machine. However, when the air fills with fog, dust, or smoke, or when the lighting fails, these optical sensors often go blind. To navigate these difficult environments, engineers turn to millimeter-wave radar. Unlike light, radar waves can punch through visual obstructions, bouncing off walls and objects to provide a rough map of the surroundings. Yet, radar has a significant blind spot of its own: while it is excellent at measuring how fast the robot is moving toward or away from objects, it struggles to tell the robot which way it is facing. Without a reliable sense of direction, a robot can drift off course, slowly accumulating errors that eventually lead it to believe it is in a completely different location than it actually is.

A team of researchers at the Skolkovo Institute of Science and Technology has developed a new system called LoCal-RIO to solve this specific problem. Their work focuses on fusing radar data with measurements from an Inertial Measurement Unit (IMU), a small sensor package containing accelerometers and gyroscopes that track motion. The core challenge they addressed is that standard methods for combining these sensors often fail over long distances because the radar cannot correct the gyroscopes' tendency to drift. The researchers found that by separating the robot's immediate movement tracking from its long-term map building, and by using a clever method to fix the sensor's internal errors only when the robot completes a full loop, they could keep the robot on course with high precision, even in large, confusing indoor spaces like underground garages and multi-story libraries.

The system operates on two distinct levels that work together like a pilot and a navigator. The first level, the navigation graph, runs constantly and quickly, fusing the radar's speed measurements with the IMU's motion data to produce a smooth, real-time path for the robot to follow. This part of the system is designed to be fast and responsive, ensuring the robot can react to its environment instantly. However, because the radar cannot perfectly determine the robot's heading, this fast-tracking layer inevitably accumulates small errors over time, causing the robot's internal map to slowly drift away from reality. The second level, the mapping graph, works in the background. It takes snapshots of the robot's position at key moments and builds a global map of the environment. This layer is slower but more thorough, capable of recognizing when the robot has returned to a place it has visited before.

The breakthrough in LoCal-RIO lies in how these two levels communicate. In many previous systems, when the background mapper realized the robot had drifted, it would simply snap the robot's position back to the correct location on the map. While this fixed the map, it created a sudden, jarring jump in the robot's internal sense of where it was, which could confuse the control systems trying to steer the vehicle. The researchers realized that the true source of the drift was not the robot's position, but a tiny, persistent error in the gyroscope's reading. Instead of forcing the robot's position to jump, their system uses the moment of a "loop closure"—when the robot recognizes it has returned to a previous spot—to calculate exactly how much the gyroscope has been lying. It then quietly adjusts the sensor's internal bias, correcting the root cause of the drift without ever disturbing the robot's current position or velocity. This allows the robot to continue moving smoothly while its internal compass gradually becomes more accurate.

To make this work, the team had to overcome the messy nature of radar data, which is often filled with false echoes and noise. They developed a filtering technique that prioritizes strong, clear reflections from solid objects while ignoring weaker, confusing signals that bounce off multiple surfaces. This ensures that the map being built is sharp and reliable. Furthermore, the system includes a rigorous safety check before accepting any correction. It does not trust a single loop closure immediately; instead, it waits until the robot has completed several more loops and confirmed that the new path is consistent with the old one. Only then does it apply the correction to the sensor's bias, ensuring that the robot never learns from a mistake.

The researchers tested their system in a variety of challenging real-world environments, including a large underground garage, a library with wooden shelves, and a campus building with complex, curved corridors. In these tests, the system successfully mapped thousands of meters of travel, maintaining a high degree of accuracy where other radar-based systems would have failed. In one specific test involving a total loop distance of over 2,200 meters, the system kept the robot's path aligned with the true geometry of the building, demonstrating that it could handle repetitive patterns and low-reflectivity surfaces that usually confuse robots. The results showed that by focusing on calibrating the sensor's internal errors rather than just fixing the map, the robot could navigate with a level of precision that rivals more expensive and fragile laser-based systems, but with the robustness needed to operate in the foggy, dusty, and dark environments where other technologies fail.

This approach represents a significant step forward in making autonomous robots truly reliable in the real world. By treating the sensor's drift as a correctable error rather than an unavoidable flaw, and by ensuring that corrections are applied smoothly and safely, LoCal-RIO offers a path toward robots that can explore and operate in the most difficult conditions without losing their way. The system proves that with the right mathematical framework, the limitations of radar can be turned into a strength, allowing machines to see clearly even when the world around them is obscured.