Three mechanisms are at work. The first is geometric degeneracy. Scan matching aligns each new scan to a map or to the previous scan, and in a corridor with featureless walls the alignment cost is flat along the corridor axis. Censi (2007) showed how the covariance of an iterative closest point estimate can be computed in closed form, which makes the weak direction visible, and Zhang et al. (2016) proposed detecting degeneracy from the eigenvalues of the optimisation problem and updating the state only in well-conditioned directions. Pipelines that lack such a safeguard report confident and wrong motion.

The second mechanism is perceptual aliasing: different places that produce the same sensor reading. Racking is modular by design, and one bay matches the next in a laser scan and often in an image. Place recognition then proposes loop closures between distinct locations, and a single accepted false closure can corrupt a pose graph beyond repair (Cadena et al., 2016). The third is change. Pallets arrive and leave, so the occupied volume of a rack face differs between the mapping run and the working run, and people and forklifts occlude the view.

The reference datasets of the field were recorded for other purposes. The TUM RGB-D benchmark covers handheld and robot-mounted cameras in offices and a hall (Sturm et al., 2012), KITTI covers road driving (Geiger et al., 2012) and EuRoC a micro aerial vehicle in a machine hall and a small room (Burri et al., 2016). Their lasting contribution is methodological. Sturm et al. (2012) defined the absolute trajectory error, computed after rigid alignment of estimate and reference by the closed-form method of Umeyama (1991), and the relative pose error over a fixed interval. Kümmerle et al. (2009) argued for measuring errors in relative displacements between selected poses, which avoids dependence on a global frame. Zhang and Scaramuzza (2018) showed that the choice of alignment changes the reported error and must be stated. Open tools implement these measures (Grupp, 2017).

"Feature-poor" must be made measurable before it can be benchmarked. Each test aisle is described by its length, its width, the ratio of the two, the type of rack face (open beams, mesh, solid panels or wrapped pallets), the fill level at recording time, the floor finish and the lighting. From these the protocol defines three classes: a reference class of short aisles with open, varied rack faces; a degenerate class of long aisles with uniform or solid faces; and an aliased class in which several parallel aisles are identical to one another. Reporting results by class prevents a good average from hiding failure in the class that matters.

All pipelines must be fed from one recording, so the rig carries the union of the sensors they need: a planar safety-type laser scanner at the usual mounting height, a three-dimensional lidar, a stereo or depth camera, an inertial measurement unit and wheel encoders. Sensors are hardware-synchronised where possible and time offsets are estimated otherwise. Intrinsic and extrinsic calibration follows published practice for joint spatial and temporal calibration of cameras and inertial sensors (Furgale et al., 2013), and the calibration files and their residuals are released with the data, since a poor extrinsic estimate is easily mistaken for a weakness of the algorithm.

Motion capture, the reference used in small indoor datasets, cannot cover an aisle of industrial length, and satellite positioning is unavailable under a roof. The protocol therefore uses surveyed fiducials. Visual tags of the AprilTag family (Olson, 2011) are fixed to rack uprights at regular spacing on both sides, and their positions are measured with a total station from stations tied into a common control network. A dedicated upward- or side-facing camera, not available to the pipelines under test, observes the tags and yields reference poses wherever a tag is in view. The uncertainty of the reference is estimated from the survey and stated; errors below it are not interpreted. Tags are masked from the images supplied to visual pipelines so that the reference does not become a feature.

Each trajectory is chosen to provoke one mechanism. A straight traverse of a single long aisle at constant speed isolates along-aisle drift. A serpentine route through adjacent identical aisles invites false loop closures. A route that leaves the racking, crosses an open dock area and returns tests recovery and genuine loop closure. A repeat of the first route after the rack contents have changed tests sensitivity to change, and a run with pedestrians and trucks present tests dynamic clutter. Runs are recorded at the working speed of the vehicle, including the stops and turns in place that real missions contain. Raw data are stored unfiltered.

The pipelines compared should span the main designs in use: planar lidar SLAM by particle filter (Grisetti et al., 2007) and by graph optimisation with submaps (Hess et al., 2016), three-dimensional lidar odometry with and without tight inertial coupling (Zhang and Singh, 2014; Shan et al., 2020), and visual and visual-inertial systems (Qin et al., 2018; Campos et al., 2021). Two configuration rules apply. Each pipeline is run once with its published default parameters and once with parameters tuned on a held-out aisle that is never used for scoring, and both results are reported. Each sequence is processed several times, because pipelines with multithreading or random sampling are not deterministic.

Table 1 lists the metrics. The departure from common practice is the decomposition of error into an along-aisle and a cross-aisle component in the frame of the racking, since a single root-mean-square figure conceals the degeneracy the benchmark exists to expose.

Reports give the distribution of each metric over repeated runs, by aisle class, together with every run that failed. A pipeline that loses tracking is recorded as a failure for that run and is not dropped from the averages without comment.