LiDAR builds a point cloud of the world — millions of range measurements per second, each one a dot in three-dimensional space. To estimate motion from consecutive clouds, odometry algorithms must match points between frames. The matching depends on voxelization: dividing space into small cubes and working with representative points. The problem is that the right voxel size indoors is wrong outdoors. A hallway needs fine resolution. A parking lot needs coarse resolution. A robot walking through a door needs to change resolution mid-stride.
Lee et al. solve this with a PID controller — the same feedback mechanism that governs thermostats and cruise control. Their system, GenZ-LIO, monitors the density of points per voxel in real time. When density drops (open space, fewer surfaces), the controller shrinks voxels to maintain sufficient geometric detail. When density spikes (cluttered interior), it enlarges voxels to avoid computational overload. The target is not a specific voxel size but a consistent point density: the system regulates the downstream quantity it actually needs rather than the upstream parameter it happens to control.
They add two more components — a hybrid metric combining point-to-plane and point-to-point matching for geometrically degenerate scenes, and a voxel-pruning strategy that eliminates unlikely correspondence candidates. The system works across indoor, outdoor, and transitional environments without manual tuning.
Most adaptive systems tune their own settings — step sizes, filter bandwidths, resolution parameters. But the parameter being tuned is usually a proxy for the quantity that actually matters. Voxel size is a proxy for point density. Learning rate is a proxy for gradient signal-to-noise ratio. When you close the feedback loop around the derived quantity instead of the proxy, the system becomes invariant to the environmental changes that make the proxy unreliable. Regulate what you need, not what you control.