For a robot to navigate the world, it must first understand itself. It needs to know exactly where its "eyes" (cameras or LiDAR) are positioned relative to its "inner ear" (Inertial Measurement Unit, or IMU). This spatial relationship, known as extrinsic calibration, is rarely static; sensors shift, mounts vibrate, and thermal expansion alters alignments.
While traditional calibration requires manual intervention with checkerboards, a growing body of research focuses on motion excitation: the idea that a robot can "learn" its own physical configuration simply by moving in specific, mathematically rich ways.
The Problem of Observability
The fundamental challenge in self-calibration is "observability." In many common scenarios, the parameters a robot needs to estimate are mathematically indistinguishable from one another. For example, as Hong Liu and colleagues explain in Online Initialization and Extrinsic Spatial-Temporal Calibration for Monocular Visual-Inertial Odometry, a visual-inertial system cannot easily tell the difference between the constant pull of gravity and the internal bias of an accelerometer without sufficient motion.
If a robot moves only in a straight line at a constant speed, the equations governing its state have infinite solutions. To "observe" the hidden variables—like the 6-degree-of-freedom offset between a LiDAR and an IMU—the robot must experience "excitation." As Srikanth Saripalli demonstrates in Motion based Extrinsic Calibration of a 3D Lidar and an IMU, successful calibration requires motion that excites at least two degrees of rotational freedom. Simple linear trajectories are insufficient to resolve the full spatial relationship.
Breaking Symmetries through Diverse Maneuvers
Why does complex motion solve the problem? It breaks the geometric symmetries that keep parameters hidden. Weikun Zhen and the team at Carnegie Mellon found in A Joint Optimization Approach of LiDAR-Camera Fusion for Accurate Dense 3D Reconstructions that achieving millimeter-level accuracy requires movement across multiple stations with non-collinear rotation axes.
This principle applies across sensor types:
- Cameras: Arno Solin shows in Robust Gyroscope-Aided Camera Self-Calibration that varied motion is required to make lens distortion and focal length observable.
- Human Motion Capture: Even in wearable tech, Chengxu Zuo notes in Transformer IMU Calibrator that a lack of rotational variety prevents systems from updating sensor offsets as they shift on the body.
- Legged Robots: Patrick M. Wensing argues that certain inertial parameters of a robot's limbs only become observable during "free-body" motion, rather than when the robot is fixed to a base.
Active Excitation: Planning to Learn
If specific motions are required for calibration, why wait for them to happen by chance? Researchers like Karol Hausman and Stephan Weiss 0002 have pioneered observability-aware trajectory optimization.
In their work, Observability-Aware Trajectory Optimization for Self-Calibration With Application to UAVs, they propose that a drone should purposefully "wiggle" or oscillate as it flies toward a goal. These intentional maneuvers maximize the information gain for the system's internal filters. Similarly, James A Preiss and Gaurav S. Sukhatme showed in Trajectory Optimization for Self-Calibration and Navigation that these "excitation maneuvers" allow a quadrotor to achieve twice the positioning accuracy of a robot following a standard, smooth, energy-efficient path.
The Limits of Motion-Based Calibration
The reliance on motion creates a paradox: the system is most blind when it is most stable. In Wide-Baseline Multi-Camera Calibration Using Person Re-Identification, Yan Xu and Kris Kitani highlight that methods using moving subjects as "dynamic keypoints" fail entirely in static environments.
This is echoed by Haojie Ren in TrajMatch, where roadside LiDARs cannot synchronize if the road is empty. Without the "time-varying features" provided by movement, the mathematical constraints necessary to link sensors across space and time simply do not exist. For autonomous systems, movement is not just a means of travel—it is a fundamental requirement for self-awareness.