When to Trust the BIM: Identifiability-Aware LiDAR-Inertial--BIM Fusion under As-Built Deviations

Yujie Zhang; Yuxuan Guo; Xinbao Sun; Chunlei Zhao; Langqi Miao; Xiaohui Jia; Jinyue Liu

cs.RONov 6, 2026

When to Trust the BIM: Identifiability-Aware LiDAR-Inertial--BIM Fusion under As-Built Deviations

Authors: Yujie Zhang, Yuxuan Guo, Xinbao Sun, Chunlei Zhao, Langqi Miao, Xiaohui Jia, Jinyue Liu

Organizations: School of Mechanical Engineering, Hebei University of Technology, No. 5340 Xiping Road, Beichen District, Tianjin, 300401, China

Abstract

Building information models (BIM) can provide a global geometric reference for robot localization with LiDAR-inertial odometry (LIO). However, as-built deviations from the as-designed model mean that a disagreement between a scan and the BIM may arise from a pose error, an offset wall surface (an as-built deviation or an unmodeled object in front of the wall), or a shifted room, and in some geometries these explanations produce identical observations. In this paper, we formalize this attribution problem as fault identification over three fault classes and derive a geometric criterion for when attribution is possible. Building on it, we propose an identifiability-aware method that reports an attribution only when the fault is attributable; otherwise, it retains the candidates, defers BIM updates and BIM-based pose corrections in the ambiguous directions, and recovers once new references arrive. In the pre-registered two-dimensional study, on cases whose fault is detectable but not attributable, the method's conditional error rate was 0.83%, against 31.67% for a baseline with uncertainty-based abstention at the same answer rate. On the 5,600 attributable cases, the method reported a unique attribution 5,574 times, with no incorrect attributions. In paired fault injections in a laboratory room with its IFC model and on three floors of the SLABIM dataset, it raised no off-target attribution (0 of 168 conditions in the laboratory; 0 of 1,302 on SLABIM) and attributed all eight laboratory faults predicted to be attributable, whereas both reproduced baselines raised off-target structural flags in every setting.

Figures & tables

Figure 1: Paired worlds. (a) World A: the as-built room is offset by

d=0.20

m from its BIM position and the LiDAR-inertial odometry (LIO) pose is correct. (b) World B: the room matches the BIM and the LIO pose is offset by

d

. (c) Both worlds give the same scan and the same LIO pose, so no method can be right in both.

Work	What it estimates	Tests, per conflict, whether the cause is attributable	Output when the cause cannot be told apart	Role in this paper
diS-Graphs [ 36 ]	Global alignment and wall and room deviations	✗; staged estimation starting from walls without deviation	One estimate of pose and deviations	Main baseline (reproduced)
BIM-Loc [ 54 ]	Per-patch structural consistency that gates BIM constraints	✗; per-patch gating, no offset size or attribution	Localization continues with the gated structure	Real-data baseline (reproduced)
LIO-BIM [ 39 ]	Pose; BIM factor added only after a match passes thresholds	✗; match-quality thresholds	Accepts or rejects the match by its quality; one pose either way	Gate reproduced
Floor-plan and BIM localization [ 9 , 28 , 20 , 49 , 21 , 37 ]	Pose, with the plan or BIM as the map	✗; iS-Graphs waits for a unique match in global localization	One pose, with the model as the reference	Background
BIM update and change detection [ 38 , 44 , 45 , 46 ]	Component existence, elements missing from the model, object-level as-built update	✗	Updates or flags the model	Background
Scan-vs-BIM [ 11 , 4 , 31 ]	Component recognition, deviation and quality	✗; error sources identified offline from deviation patterns [ 4 ]	–	Background

Table 1: Comparison with the prior work on scan–BIM conflict attribution. ✗: no; ✓: yes.

Figure 2: The three fault classes of the single-fault model, each shown at 0.20 m. (a) LIO bias: the LIO pose is offset in

x

y

and yaw (three parameters). (b) Wall-surface offset: the surface observed at one BIM wall is offset along the wall normal by

\delta

(one parameter). It may come from an as-built deviation of the wall (b1) or from an unmodeled object in front of it (b2); the wall plane alone cannot separate the two, and the method does not distinguish them. (c) Room translation: the whole room is translated in the plane (two parameters) relative to the datum, the start coordinates propagated through the window priors. Normal construction tolerance is treated as noise, not as a fault.

Figure 3: (a) Single-room view: an aligned scan raises no conflict; a translation is explained by a room translation and by an LIO bias alike; a rotation is explained only by the LIO bias, which adds yaw. Anything a room translation explains, an LIO bias explains as well, so a room translation can never be confirmed on its own. (b) Witness ladder, Manhattan special case. Each line is one witness along the offset direction, and its dot marks the robot position that witness implies. With one witness, a fault produces no contradiction; with two, it is detected but cannot be attributed; with three or more, it can be attributed.

Outcome	Working pose	BIM
No conflict	Fit of H0 with all walls; directions that the BIM cannot verify are marked as unverified	Unchanged
Unique attribution	Corrected according to the attributed explanation	Corrected copy proposed for verification; the original is not replaced
Ambiguous	Correction of the lowest-cost remaining explanation in the directions where all candidates agree; prior and odometry in the ambiguous directions	Unchanged
All explanations excluded	Position withdrawn and reported as unavailable	Unchanged

Table 2: Working pose and BIM output for each outcome of the tests.

Figure 4: Pipeline of the identifiability-aware LiDAR-inertial–BIM fusion. LiDAR-inertial odometry runs without the BIM; the BIM enters through wall-patch association in fixed-length path windows. All single-fault explanations (H0, each visible wall, each visible room and an LIO bias) are fitted, a conflict is detected against H0, and explanations are excluded pairwise. If one explanation remains, it is attributed and its BIM change is only proposed for verification. If several remain, they are retained and BIM anchoring is deferred in the directions where they do not agree on a BIM-anchored position. If none remains, the position is withdrawn. The pose interval is the union over compatible explanations, an ambiguous axis is withdrawn after a preset distance, and a new reference triggers re-identification.

Data	BIM	Datum	Truth or reference	Use
Two-dimensional generator	Corridor, corner, room, room with pillar; fixed references given	Given prior (strong, weak or none)	Known truth	Pre-registered study (comparison, recovery and paired-world grids); deferred-anchoring demonstration; fixed-window recovery test (RQ1–RQ4)
Laboratory room 209 (209_1, 209_2, 209_3)	IFC model of the room	Laser-tracker initial pose for the test sequences 209_2 and 209_3; 209_1 calibrates the noise and the tracker to the IFC	Initial pose only	Paired injections (RQ5)
SLABIM 1F (sessions 1F2, 1F3)	Mesh BIM	BIM alignment of the first nine keyframes	Per-frame reference; about 18 cm disagreement between sessions	Development floor: full sessions on the original BIM, paired injections (RQ4, RQ5)
SLABIM 3F and 4F	Mesh BIM	BIM alignment of the first nine keyframes	Per-frame reference recovered from the map	Paired injections, accessed after the front-end freeze (RQ4, RQ5)

Table 3: Data, datum, truth and use of each data source.

Figure 5: Conditional error rate against answer rate on the 7,200 detectable-but-not-attributable cases of the pre-registered two-dimensional study. The line is the diS-Graphs-style baseline with uncertainty-based abstention; markers are native operating points. At the answer rate of Ours (120 of 7,200, 1.67%), Ours made one wrong attribution (0.83%) and the abstaining baseline 38 (31.67%); the paired risk difference is

-30.8

percentage points (99.58% CI

-44.5

-19.3

). Restricted to wall faults, the rates are 0.84% and 31.09%. Error rates are cell macro-averages.

Label	Faults	Ours	Forced variant	diS-Graphs-style
Detectable but not attributable	All ( $N=7{,}200$ )	$A=120$ (1.67%); $E/A=0.83\%$	$A=7{,}196$ ; $E/A=51.00\%$	Native: $A=7{,}196$ , $E/A=78.11\%$ ; abstaining at the answer rate of Ours: $E/A=31.67\%$
	Walls ( $N=2{,}400$ )	$A=119$ ; $E/A=0.84\%$	–	Abstaining at the same answer rate: $E/A=31.09\%$
Attributable	All ( $N=5{,}600$ )	$A=5{,}574$ ; no wrong attribution	$A=5{,}595$ ; no wrong attribution	$A=5{,}596$ ; 2,705 correct (48.30%)
	Walls ( $N=3{,}200$ )	3,200 correct	–	2,705 correct (84.53%)
	LIO ( $N=2{,}400$ )	2,374 correct	–	0 correct (no LIO output)
Undetectable	All ( $N=16{,}000$ )	$A=164$ ; $E/A=0.61\%$	$A=1{,}854$ ; $E/A=44.50\%$	$A=11{,}443$ ; $E/A=94.49\%$

Table 4: Pre-registered two-dimensional study by label and fault type (strong and weak priors, translations larger than 0.05 m).

A

: answers;

E/A

: conditional error rate (cell macro-average). Counts are unweighted.

Label	$N$	Missed	Wrong primary	Extra only	Correct	Strict	Lenient
Attributable	5,600	4	2,681	210	2,705	48.30%	52.05%
Det. not attr.	7,200	4	5,426	195	1,575	21.88%	24.58%
Undetectable	16,000	4,557	10,662	151	630	3.94%	4.88%
Flags under an LIO fault			Runs		Share
Walls and rooms			9,807		68.10%
Rooms only			86		0.60%

Table 5: Errors of the diS-Graphs-style baseline in the pre-registered two-dimensional study. Top: outcome of each run by label; strict scoring requires additional flags to meet the 3 cm rule, lenient scoring ignores them. Bottom: flags raised under an LIO fault (14,400 runs) and the magnitude of the flagged offsets (per flagged element).

Figure 6: Deferred BIM anchoring in the paired worlds with a strong prior (

\sigma=0.03

m; 200 trials per world). (a) Working-pose error along the ambiguous direction in the room world (a1) and the LIO world (a2): median and 10–90% range. Before the reference becomes visible, the pose follows odometry in the ambiguous direction, so the LIO world carries the prior offset (RMSE about 0.21 m, against 0.03 m in the room world); once the reference is visible, both errors fall to about zero. (b) Share of trials in each output state. Ambiguity was resolved correctly in 200 of 200 room-world and 194 of 200 LIO-world trials, with no wrong unique attribution.

Figure 7: Resolution of ambiguity after a new reference appears (200 trials per configuration and world). (a) Resolution distance of each trial, measured from first detection. (b) Kaplan–Meier estimate [ 26 ] of the share still unresolved against distance travelled in the room world (b1) and the LIO world (b2); ticks mark censoring and the dashed line marks the 10 m distance cap. With the reference at 16 m, all 200 trials in each world are first withdrawn at the cap; the ambiguity is then resolved in 200 room-world and 195 LIO-world trials. Without a reference, every trial is withdrawn after 10 m. With the weak prior, the LIO offset went undetected in 196 LIO-world trials, where H0 is already correct; these are not in the risk set.

Figure 8: Laboratory room 209 with its IFC model; both test sequences start from the laser-tracker initial pose. (a) Point cloud over the IFC walls, with the trajectories, the four windows and the injection targets. (b) Native outputs by predicted label, and off-target attributions or structural flags by fault type, over 168 injected conditions. Ours attributed all 8 predicted-attributable faults correctly and raised no off-target attribution; the diS-Graphs-style and BIM-Loc-style baselines raised off-target structural flags in 63 and 32 conditions. Of the 14 detectable-but-not-attributable conditions, Ours attributed 8 correctly, retained candidates in 4 and found no conflict in 2; four of these six ended with the position unavailable. The diS-Graphs-style baseline flagged 7–14 structures in each of these 14 conditions, none of them a new off-target flag relative to its C0. (c) Median absorption by fault type on the conditions shared with Ours. Trust BIM and the LIO-BIM-style gate report no attributions or structural flags, so they appear only in (c). In a single room, a room translation and an LIO translation form paired worlds, so every method that corrects the pose with the BIM follows the room translation (0.955–0.998). Ours does not misreport it, and flags $x$ and $y$ as unverified in 23 of its 26 control outputs. The four windows come from two sequences, so the counts are descriptive.

Figure 9: Ours on the original BIM over two full sessions of SLABIM floor 1F, the development floor. (a) Session 1F2 (148 outputs) and (b) session 1F3 (106 outputs), with keyframes colored by output state; unique attributions cannot be verified on the original BIM. (c) Keyframe 9 of 1F3: the two observed surfaces associated with walls 619 and 6459 are 1.776 m apart, against 2.625 m in the BIM. Where the scan cannot be reconciled with the BIM, Ours withdraws the position rather than forcing an answer. The recovered references disagree by about 18 cm between the sessions, so the states are descriptive.

Figure 10: The 1F3 segment where Ours withdraws its position (keyframes 9–16, about 18–32 m, 890 frames). (a) Translation error against the recovered reference: the BIM-Loc-style backend (RMSE 0.395 m) is worse than FAST-LIO2 alone (0.118 m) in every frame. (b) Absolute error and self-reported 95% half-width per axis, log scale. The median per-axis

\sigma

is about 0.56 mm, and the nominal 95% intervals include the reference in 0% (

x

z

) and 2.7% (

y

) of the frames. The reproduction uses an odometry prior 8–10 times tighter than measured. A covariance computed on the assumption that the map is correct cannot reflect errors in the map.

			Off-target			Absorption (%)					Reference inclusion
Floor	Fault	$N$	Ours	diS	BIM-Loc	Ours	Trust BIM	diS	BIM-Loc	Gate	Ours	Trust BIM	BIM-Loc
Room 209	Wall offset	48	0	0	7	6.7 (42)	17.1 (42)	-0.1	0.3	0.0	–	–	–
	Room or region translation	48	0	0	0	99.8	99.8	–	95.5	0.0	–	–	–
	LIO prior translation	48	0	48	17	0.3	0.3	100.0	6.8	100.0	–	–	–
	LIO prior yaw	24	0	15	8	0.0	0.0	19.7	1.5	100.0	–	–	–
1F2 (dev.)	Wall offset	18	0	0	6	78.1	78.2	1.5	0.0	0.0	18/18	8/18	0/18

Table 6: Paired injections on the real data, by floor and fault type (

N

conditions). Off-target: conditions with an off-target attribution (Ours) or off-target structural flag (baselines), new relative to each method’s own C0; a condition that a baseline does not support counts as having no flag, and the diS-Graphs-style baseline supports no room translation in room 209. Absorption: median share, in %, of the injected offset that the pose follows; the trust-BIM ablation is evaluated on the conditions where Ours has a value, the other methods on their own valid conditions, with the number of conditions in parentheses when it is smaller than

N

. Reference inclusion: conditions whose interval includes the recovered reference; not available in room 209. Room 209 pools its two test sequences, 209_2 (one window) and 209_3 (three windows). Region translations on 4F move a corridor region. 1F is the development floor.

Figure 11: Paired fault injections on SLABIM floors 1F, 3F and 4F (1F is the development floor). (a) Conditions with an off-target attribution (Ours) or off-target structural flag (baselines), counted as new relative to each method’s own C0, by floor. Ours raised none in 1,302 conditions; the two baselines raised 391 and 1,070. (b) Native outputs by predicted label; BIM-Loc-style flags are not causal attributions. (c) Median absorption of the injected offset into the pose for the trust-BIM ablation and the working pose of Ours, on the conditions where both are available. Region injections are geometrically degenerate and form paired worlds, so both absorb them almost fully. Window blocks are few, so the results are descriptive.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 1: Four layouts of the two-dimensional simulation study. (a) Corridor (width 2.4 m). (b) Corner. (c) Room with a door (6

\times

4 m). (d) Room with a pillar. Black squares and thick black segments are fixed references; the blue line is the trajectory (keyframe every 2 m); the dashed segment indicates no later reference.

Label	Method	$N$	$A$	$E$	$E/N$ micro	$E/N$ macro	$E/A$ macro	Coverage macro
Undetectable	Ours	16,000	164	1	0.006%	0.006%	0.610%	98.78%
	Forced variant	16,000	1,854	825	5.156%	5.156%	44.498%	94.46%
	diS-Graphs-style	16,000	11,443	10,813	67.581%	67.581%	94.494%	84.08%
	diS-Graphs-style, first reconstruction	16,000	15,400	15,055	94.094%	94.094%	97.760%	81.69%
Detectable but	Ours	7,200	120	1	0.014%	0.014%	0.833%	95.14%
not attributable	Forced variant	7,200	7,196	3,670	50.972%	50.972%	51.001%	46.90%

Appendix

Table 1: Main table of the pre-registered two-dimensional study (strong and weak priors, translations larger than 0.05 m).

N

A

and

E

are unweighted counts; macro-averages weight the cells equally. The rows of the first reconstruction are descriptive and not part of the pre-registered comparison.

	Det. not attr. ( $N=11{,}200$ )			Attr. ( $N=8{,}000$ )	No fault
Variant	$A$	$E/A$	Coverage	Answer rate	Flags
Full method	1,227	0.08%	95.10%	92.12%	0/4,800
Forced variant	10,428	44.36%	53.58%	92.39%	0/4,800
Without LIO	4,905	78.02%	75.34%	62.45%	0/4,800
Without room	6,046	39.28%	74.69%	92.16%	0/4,800

Appendix

Table 2: Suspect-set ablation over the whole main grid of the pre-registered study (native operating points; rates are cell macro-averages).

Grid	Method	Never detected	Detected but unresolved	Wrong attribution	Correct and stable	Median distance to resolution (m)
Recovery	Ours	29.94%	27.42%	0.42%	42.22%	8.0
	Forced variant	29.94%	0.28%	27.58%	42.19%	0
	diS-Graphs-style	4.16%	0.00%	75.38%	20.46%	not reached
	diS-Graphs-style, first reconstruction	0.07%	0.00%	97.92%	2.01%	not reached
Paired worlds	Ours	43.42%	22.83%	0.19%	33.55%	8.0
	Forced variant	43.42%	0.95%	20.86%	34.76%	0

Appendix

Table 3: Outcomes on the recovery grid (14,400 runs) and the paired-world grid (19,200 worlds) of the pre-registered study. A run is a wrong attribution if a wrong unique attribution ever occurred; it is correct and stable if every keyframe after the first correct attribution stays correct. The median distance to resolution is the Kaplan–Meier median [ 26 ] from first detection, with unresolved runs right-censored; the forced variant answers on detection, so its median is 0, and the diS-Graphs-style baseline, which has no separate alarm, takes its first answer as detection. Half of the cells in each grid have no later reference and cannot be resolved; the two worlds of a pair are not independent. The analysis is descriptive and not a pre-registered criterion.

Parameter	Value
Shared by both configurations
Detection level $\alpha$	0.01
Exclusion level per ordered pair	$0.01/m^{2}$
Bound on a wall-surface offset	0.35 m
Offset above which a wall or room attribution is proposed as a BIM revision	0.05 m
Nominal level of the pose intervals (per projection)	0.95

Appendix

Table 4: Shared method parameters and parameters of the simulated configuration.

Parameter	Value
Keyframe spacing	2 m
Window length; step between windows	9 keyframes; 8 keyframes
Prior standard deviation, translation (each axis)	0.289 m
Prior standard deviation, yaw	0.101 rad
Range noise $\sigma_{r}$	0.0836 m
Shared wall tolerance $\sigma_{t}$	0.04 m

Appendix

Table 1: Parameters of the real-data configuration.

LiDAR Localization

LiDAR (light detection and ranging) localization estimates a robot or vehicle’s position and orientation from laser scans and maps.

See all 43 papers

Jun 12, 2026

BIM-Loc: BIM-Integrated Discrepancy-Aware LiDAR-based Indoor Localization

Accurate and robust localization is a fundamental requirement for service and inspection robots, particularly in feature-sparse indoor environments where traditional systems struggle due to a lack of distinct landmarks. While prior maps can enhance robustness, precise and compact maps capturing real-world details are often unavailable for new or frequently changing environments. This paper presents BIM-Loc, a novel discrepancy-aware LiDAR-based localization method that directly integrates Building Information Models (BIM) from the design phase. BIM-Loc simultaneously estimates trajectories aligned with the BIM coordinate system and identifies discrepancies between real-world observations and the as-designed BIM in an online fashion. Our core contributions include: (1) a novel multi-hit ray casting strategy for efficient BIM-point data association and projection of 3D observations into 2D texture space; (2) a pose graph optimization framework with BIM-integrated factors that enforces consistency among odometry, sequential scans, and BIM structures; and (3) a hierarchical Bayesian inference module that incrementally updates a continuous 2D surface representation for discrepancy detection, propagating updates from the pixel to the structure level. Extensive evaluations in both simulation and real-world applications demonstrate that BIM-Loc significantly outperforms state-of-the-art map-based methods in localization accuracy and robustness.

Yinqiang Zhang, Liang Lu, Yipeng Pan +5

Department of AI & Data Science, University of Hong Kong (HKU), Hong Kong, China. · Humanoids and Human Centered Mechatronics Research Line, Italian Institute of Technology (IIT) Genoa, Italy. · Department of Architecture and Civil Engineering, City University of Hong Kong (CityU), Hong Kong, China.

Pose Graph Optimization LiDAR Localization

Ask PDF

Jul 19, 2026

Move First, Commit Later: Selective LiDAR-to-BIM Global Initialization by Verifying Proposals Against Competing Explanations

Global LiDAR-to-BIM initialization must place a robot within an as-designed building model without a prior pose. In repetitive interiors, the principal failure mode is not low-confidence registration but confident aliasing: a submap can match several BIM regions with comparable scores, producing a high-scoring pose displaced by symmetry. We present Move First, Commit Later, a selective layer that treats a registration front-end as an evidence source and decides whether to commit. Candidates from multiple submaps are mapped to a common

\mathrm{SE}(2)

anchor; a top-1 consensus

B_m

, invariant to non-champion multiplicity, aggregates cross-submap evidence; and topology serves only as a binary feasibility gate. The decision is typed--COMMIT, DEFER, or AMBIGUOUS(

τ

), reporting the detected symmetry period--and reversible: symmetry-breaking motion upgrades AMBIGUOUS to COMMIT. On a real multi-room building and a controlled symmetric simulation, the layer commits correctly in every tested trajectory-scale configuration and otherwise abstains with a typed state, whereas forced-choice policies on the same front-end select wrong rooms in most cases. Committed anchors are within 0.02--0.36 m of an independent laser-tracker position reference and within

1.9^\circ

of a BIM-registration orientation proxy. The evaluation covers one building and one front-end; the layer is designed to be front-end modular.

Yujie Zhang, Yuxuan Guo, Zeheng Fan +3

School of Mechanical Engineering, Hebei University of Technology, Tianjin 300401, China.

Point Cloud Registration LiDAR Localization Map-Based Localization

Ask PDF

Sep 17, 2025

BIM Informed Visual SLAM for Construction Environments

Monitoring building construction sites requires comparing the as-planned design with the as-built state, which can be estimated in real time using Simultaneous Localization and Mapping (SLAM) techniques. However, visual SLAM is prone to trajectory drift in construction environments, producing maps that are geometrically inaccurate with the actual environment. To address this limitation, we augment an existing RGB-D SLAM system with structural priors derived from the Building Information Model (BIM). The system associates detected walls with their BIM counterparts and includes these correspondences as geometric constraints in the back-end optimization, reducing drift and enhancing global consistency. The proposed method operates in real time and is validated on multiple real construction sites, achieving an average trajectory error reduction of 25.23% and a 7.14% improvement in map accuracy over state-of-the-art baselines. Robustness analyses further demonstrate resilience to incomplete BIM data and geometric discrepancies between as-planned models and the as-built environment.

Asier Bikandi-Noya, Miguel Fernandez-Cortizas, Muhammad Shaheer +3

Automation and Robotics Research Group, Interdisciplinary Centre for Security, Reliability, and Trust (SnT), University of Luxembourg, Luxembourg

Simultaneous Localization and Mapping Visual SLAM

Ask PDF

Abstract

LiDAR Localization

Explore similar work

BIM-Loc: BIM-Integrated Discrepancy-Aware LiDAR-based Indoor Localization

Move First, Commit Later: Selective LiDAR-to-BIM Global Initialization by Verifying Proposals Against Competing Explanations

BIM Informed Visual SLAM for Construction Environments