A robot method waits when scans cannot explain a model conflict
When to Trust the BIM: Identifiability-Aware LiDAR-Inertial--BIM Fusion under As-Built Deviations · arXiv:2611.09045
In plain words
A robot can compare what its laser sees with a building’s digital model. If they disagree, the cause might be a wrong robot position, a moved wall, or a shifted room. This paper introduces a method that checks whether the available walls and references can tell those causes apart.
When they cannot, it keeps the possible causes instead of choosing one and changing the robot’s position or model on a guess. Here, detectable means the data can indicate a fault; attributable means they can distinguish its cause from competing explanations.
In a two-dimensional simulation, it made fewer wrong attributions than a comparison method that abstained at the same answer rate on cases labeled detectable but not attributable.
The research
A scan–model mismatch can lead a robot to correct its position using a faulty model, or to revise the model based on an odometry error.
The method changes the decision: a unique cause is reported only when independent geometric references distinguish it, while unresolved causes remain candidates instead of guiding a model-based correction.
The method checks whether geometry can distinguish an odometry error, offset wall or shifted room, then reports a cause only when one explanation remains.
In a pre-registered two-dimensional simulation, researchers compared the method with a diS-Graphs-style system that abstains when its uncertainty is high. Here, detectable means the data can indicate a fault, while attributable means the evidence can distinguish its cause from competing explanations.
They matched the number of reported attributions on cases labeled detectable but not attributable, then compared wrong reports among those answers. The method’s conditional error was lower, while it also uniquely attributed most cases labeled attributable without error. This supports retaining competing causes when the available geometry cannot choose between them.
The method assumes one active fault at a time and models horizontal wall offsets only up to 0.35 m. The real-sequence tests cover one laboratory room and three floors of one university building, with limited reference data.
What the evidence shows
- A conflict has several causes
The identifiability-aware LiDAR-inertial–BIM fusion method is a robot-localization method that decides whether a disagreement between a laser scan and a building model can be assigned to an odometry error, a wall offset or a room shift.
The reason for asking is concrete: acting on the wrong explanation can pull the robot’s position toward a faulty model or support a mistaken model revision.
- The same data can disagree
The paper’s paired-world example shows the ambiguity: a room shifted from its modeled location with correct odometry can produce the same scan and odometry pose as a correctly placed room with an odometry error. The observations therefore cannot always justify a unique cause.
The method asks whether the geometry includes independent references that constrain the offset direction before it accepts an explanation.
- References decide what is knowable
Independent walls can act as witnesses because they constrain the robot’s position in different directions. A single-room translation can still fit both a room shift and an odometry error; a further reference can separate them.
Whether a particular conflict is distinguishable also depends on offset size relative to noise, so collecting more data in unchanged geometry may not help.
- Fewer wrong reports on ambiguity
In a pre-registered two-dimensional simulation, detectable means the data can indicate a fault, while attributable means the evidence can distinguish its cause from competing explanations. On cases labeled detectable but not attributable, the method’s conditional error rate was 0.83%, compared with 31.67% for an uncertainty-abstaining baseline at the same answer rate.
This measure counts wrong reports among reported attributions, not among all cases. It supports the claim that retaining competing explanations reduced errors at that operating point.
- Answers when evidence separates
The method also uniquely attributed 5,574 of 5,600 simulated cases labeled attributable without an incorrect attribution. Paired fault injections in one laboratory room and on three floors of one university building showed no off-target attributions by the method, while reproduced baselines raised off-target structural flags.
Those real-sequence results are descriptive, and the tested settings do not establish performance across buildings or every positioning fault.
Deep
- Original paper authors
- Yujie Zhang, Yuxuan Guo, Xinbao Sun, Chunlei Zhao, Langqi Miao
- Submitted to arXiv
- License
- Not specified by source