acceptodds
Under review as a conference paper at ICLR 2027

Elevator-VIGS: Separating Elevator Motion from Robot Motion in Visual-Inertial Gaussian Splatting SLAM

Abstract

We present Elevator-VIGS, a visual-inertial 3D Gaussian Splatting SLAM system that keeps tracking and mapping through elevator rides. Inside a moving elevator, the two sensors are in conflict. The camera sees only the robot's motion relative to the elevator, while the IMU senses that motion plus the elevator's motion relative to the world. This conflict is challenging for existing visual-inertial estimators. If vision dominates, the estimator tracks only the robot's motion within the elevator and misses the elevator's rise, and if the conflict remains, the estimator diverges. We observe that the conflict comes from forcing both observations into a single coordinate frame. We instead estimate the robot's pose in the elevator's coordinate frame, and the elevator's motion relative to the world as a per-keyframe transport state, the elevator's rise and vertical velocity, within dense visual-inertial bundle adjustment. Elevator-VIGS detects rides zero-shot with a vision-language model and a depth network, and constrains the transport state at the departure and the arrival. We record real-world and simulated elevator sequences. On these sequences, Elevator-VIGS achieves state-of-the-art tracking and rendering performance. On four elevator-free public benchmarks it keeps the state-of-the-art performance of VIGS-SLAM. Code and datasets will be made public.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Visual-Inertial-SLAM, Gaussian Splatting, Elevator Travel

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_YL9FOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The authors propose Elevator-VIGS, a VI-based 3DGS SLAM system for use through elevator rides. Unlike earlier works which use LiDAR or other sensors, they do so using vision and IMU. The core challenge here is that the RGB and IMU data are conflicting inside the elevator, making pose estimation challenging. They address this by using a transport state which separates elevator and robot motion, detecting elevator rides using zero-shot VLM detection and anchored by arrival and departure (0 velocity) conditions of the elevator. The system is evaluated on a wide range of both sim and real scenes, both enclosed and glass elevators, and substantially outperforms prior work, on the evaluated sequences, while generating a GS map.

Critical strengths
  • The paper addresses a challenging problem of SLAM in elevators, and to my knowledge is the first GS SLAM system to do so.
  • The paper is comprehensive and well evidenced, providing real world experiments in a variety of elevators, including glass ones.
  • The system performs well on reported sequences and is compared against a wide range of methods using different modalities.
Critical weaknesses

NIL.

Questions for authors
  • How does the system perform in glass enclosed or semi-glass elevators? The authors do report numbers for the detector, P_{elev} drops 0.91->0.82 from enclosed to glass, while only triggering on 75% of frames in the glass ones, vs every frame for enclosed. But how does the transport state handle this? In Sect. 3.2.1, it says the visual residual is written onto the elevator-frame and not discarded, so it seems like it would conflict with the inertial residuals. Yet the results suggest that system works fine in practice, so I would like to see some numbers on how the transport state performs in glass elevators, and how it handles the visual residuals in this case.
  • C.2, C.4 seem to imply the system can handle elevators which stop at multiple floors, but I don't see any sequences that test this? And how does it perform (i.e. Eq. 14) when the camera starts moving before the elevator stops at a floor, which is likely human behavior to pre-empt getting out.
  • The only mention of people in this text is the blurring of faces. Did any of the real world sequences include people in the elevator? I am curious how the system handles people or other objects in the elevator, which is the most common case in the real world? Presumably, this affects both elevator detection of the VLM and visual tracking in elevator frame.
3 comments
AuthorsOpenReview ↗

Response to Reviewer YL9F (1/2): glass elevators

Q1. Glass elevators. Glass elevators work as well as enclosed ones in tracking and in detection, and the visual and inertial residuals do not conflict. We answer the three parts of the question in turn.

Performance. The table compares the vertical error of the two elevator types on the paper's 16 real-world sequences.

SequencesElevatorCountMean vertical error (cm)Range (cm)
One-wayEnclosed6220.2 to 62.1
One-wayGlass241.7 to 7.0
Round-tripEnclosed540.1 to 16.0
Round-tripGlass390.4 to 15.6

On a one-way sequence, the robot rides to another floor, and the vertical error is the difference between the estimated height change and the laser-measured floor height HH. On a round-trip sequence, the robot rides away and returns to the starting floor, and the vertical error is the estimated height difference on returning to the start. The range is taken over the sequences of each row.

On the one-way sequences, the two glass rides have errors of −1.7 cm on Office2 (0.1% of HH) and −7.0 cm on Station1 (1.0%), below the enclosed mean of 22 cm. On the round trips, the three glass sequences Office3, Station2 and Station3 return to the start with errors of 0.4, 10.2 and 15.6 cm. Their largest error is about the largest enclosed one (16.0 cm), and the two means differ by 5 cm.

Detection on 75% of the frames. This does not hurt detection, and all five glass rides are detected. The two visual cues are not needed on every frame. They only have to fire together for 1 s to tell the detector that the robot is inside an elevator, and the departure and the arrival are then detected from the IMU. The frames on which the cues do not fire therefore cost nothing.

Visual and inertial residuals. In practice the visual residuals do not conflict with the inertial residuals, because the pixels that look through the glass receive low confidence. As in DROID-SLAM, the BA weighs each visual residual by its confidence. Once the elevator departs, the confidence of the pixels that see through the glass drops by 80 to 88%. Their visual residuals are kept, as Sec. 3.2.1 says, but they carry little weight. The pixels on the elevator itself dominate, and these pixels agree with the elevator frame.

The elevator's rise that the transport state holds comes from the IMU, so the view through the glass does not enter it.

AuthorsOpenReview ↗

Response to Reviewer YL9F (2/2): several stops, moving before the elevator stops, and people in the elevator

Q2. Several stops, and moving before the elevator stops. Both cases work with the same code and configuration as every other sequence, and every ride is detected. The paper indeed has no sequence with an intermediate stop, since its two-ride sequences are round trips, in which the robot rides to another floor, leaves the elevator and later rides back to the starting floor. We therefore recorded new sequences.

Several stops. Our formulation supports intermediate stops without any modification. Each arrival ends a ride and each departure begins a new one (App. C.4), so a trip with a stop is detected as two rides.

In two new real-world sequences, the elevator waits 11 s and 26 s at an intermediate floor with open doors and then continues. The ride detector finds both rides of each sequence and reports no false ride during the wait. It detects a departure from the IMU only when the vertical velocity stays above 0.4 m/s for 1 s, which does not happen while the elevator stands at a floor, with open doors or with passengers moving. The vertical error over the whole trip stays within 2.1% of HH.

  • New office building, HH = 14.28 m. The error is 7 cm with the stop (27 cm for a non-stop ride in the same elevator).
  • Residential2's elevator, HH = 49.97 m. The error is 102 cm with the stop (62 cm for a non-stop ride).

Moving before the stop. Our formulation also supports this case without any modification, and it hurts neither detection nor accuracy.

The robot's motion and the elevator's motion are tracked separately. The transport state carries only the elevator's vertical motion. The camera and the IMU together track the robot relative to the elevator, in rotation, horizontally and vertically. A robot that walks to the doors while the elevator decelerates is therefore tracked like any other motion inside the elevator.

Three new real-world sequences test this case. In the 4 s before the stop, the robot moves 21 to 88 cm and turns 51° to 95° (at most 16 cm and 25° in the paper's sequences). The arrival is detected in every ride. The one-way vertical errors, measured against the laser-measured floor height HH, are 0.5%, 1.9% and 2.1% of HH, within the 2.9% of the paper's one-way sequences.

Eq. 14. Eq. 14 gives the error when the robot itself moves vertically during a rest of the elevator. Our code now determines the rise from two rest windows, the few seconds in which the elevator stands still before the departure and after the stop (see our general response). The robot's velocity at the instant of the stop therefore no longer enters the rise. Only the robot's mean vertical velocity vzWv^W_z over a whole rest window still does, and it changes the rise by −vzWτ/2-v^W_z\tau/2, where τ\tau is the duration of the ride.

Q3. People in the elevator. Our system handles passengers in the elevator, and every ride with passengers is detected and tracked within the accuracy of the paper's sequences. One of the paper's sequences, Station3, has other passengers in the elevator, and we recorded three more sequences in which passengers stand and walk in front of the camera. One of the three has a mirror that reflects the passengers.

Detection. Passengers do not affect the detection of the VLM, and all five new rides are detected with no false ride reported. The median PelevP_{\text{elev}} is 0.78 to 0.95 and stays above the 0.5 threshold on every frame of the rides. The departure and the arrival are then detected from the IMU, which passengers do not affect.

Tracking. Passengers who stand still move with the elevator, so the visual residual treats them like the rest of the elevator. Walking passengers get no dedicated term. The BA weighs each visual residual by its confidence, and the pixels on walking passengers receive 3 to 5 times less confidence than the rest of the image. The inertial residuals keep constraining the robot's rotation and horizontal motion. The one-way vertical errors of the new sequences lie within the 2.9% of the paper's one-way sequences.

I appreciate the authors’ detailed responses. The additional experiments address my earlier concerns. Overall, I find the real-world value of LiDAR-free operation compelling, but certain ideas are borrowed from Elevator-LIO and applied to visual-inertial sensing. As such, I will retain my score.

Reviewer_7SUiOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper addresses accurate localization and mapping during elevator rides. The authors identify a frame conflict: the camera observes only the robot’s motion relative to the elevator, whereas the IMU also senses the elevator’s motion relative to the world. Existing visual-inertial SLAM systems may therefore miss the elevator’s rise or diverge. To address this problem, the authors propose Elevator-VIGS, which uses a transport state to represent the elevator’s rise and vertical velocity separately from the robot’s motion in the elevator frame. The system also incorporates zero-shot ride detection, departure and arrival constraints, a deferred fold, and loop closure across floors. The authors evaluate tracking and rendering against multiple SLAM baselines on real-world elevator sequences, simulated sequences with ground truth through the ride, and four elevator-free public benchmarks. They also conduct ablation and sensitivity studies.

Critical strengths
  1. The paper clearly characterizes the frame conflict in camera–IMU SLAM during elevator rides and adapts a transport-state formulation to address it within dense visual-inertial bundle adjustment. The substantial tracking gains over VIGS-SLAM make this a valuable system contribution.

  2. The experiments cover real-world and simulated elevator sequences, elevator-free benchmarks, multiple baselines, ablations, and sensitivity tests. This breadth supports the reported improvements and clarifies the role of each component

Critical weaknesses
  1. Novelty is relatively limited, with its main new contributions lying in system adaptation. Elevator-LIO already separates robot motion relative to the elevator from vertical elevator motion. Elevator-VIGS applies a similar decomposition to visual-inertial Gaussian Splatting SLAM, adding update projection, a deferred fold, ride detection, and loop closure across floors. These useful components primarily tailor the existing approach to a different sensing and mapping pipeline.
  2. The method is highly tailored to a specific elevator-motion setting, limiting its general applicability. Its formulation relies on predominantly vertical translation and identifiable departure and arrival rests. The sensitivity analysis reveals substantial degradation when these assumptions are violated, raising concerns about generalization to more varied elevator motions and operating conditions.
Questions for authors
  1. How does Elevator-VIGS compare with a direct adaptation of Elevator-LIO’s motion decomposition and rest constraints to VIGS-SLAM, and what additional benefits can be attributed specifically to update projection and the deferred fold?
  2. What role, if any, does 3DGS play in elevator-motion estimation or tracking beyond mapping and rendering, and how can its contribution be distinguished from that of the transport-state formulation?
3 comments
AuthorsOpenReview ↗

Response to Reviewer 7SUi (1/2): novelty and the direct adaptation of Elevator-LIO

W1. Novelty. We share Elevator-LIO's decomposition, but our input differs. Elevator-LIO needs a LiDAR, whereas we use only a monocular camera and an IMU. To our knowledge, no prior SLAM system handles elevator rides with these two sensors. With this input, Elevator-LIO's design does not transfer as it is. Directly adapted to our visual-inertial BA, it fails on 7 of the paper's 16 real-world sequences, and on the sequences it completes, its one-way error is 1411 cm against our 17 cm (Q1). Beyond the sensors, we therefore differ from it in the estimator, the ride detector and loop closure.

Estimator (update projection and deferred fold). Elevator-LIO runs an iterated ESKF that keeps only the current state, whereas we optimize a sliding window of keyframes in dense BA, where Elevator-LIO's fold at the arrival loses the rise (Q1). The fold is the step that adds the rise to the poses.

Ride detection. Elevator-LIO's conclusion states that its LiDAR detector "may therefore miss detections in non-standard elevators, such as fully mirrored, transparent, or large freight cabins" and proposes "visual semantics". Our ride detector takes its visual semantics from frozen pretrained models and needs no training on elevator data. All our sequences are held out for its prompts and thresholds, which we chose on recordings in three other buildings. With one configuration, it detects every ride across the elevators and conditions of our sequences, all 24 real-world rides against 15 for Elevator-LIO.

Loop closure across floors. The same conclusion names "place recognition and factor-graph optimization back-ends" as future work, which our dense BA and loop closure across floors provide.

Elevator-LIO is also an unpublished arXiv preprint, which the ICLR 2027 Reviewer Guide does not require submissions to compare against.

Q1. Direct adaptation of Elevator-LIO. The direct adaptation does not work in our visual-inertial BA. It fails on 3 of the 8 one-way and 4 of the 8 round-trip real-world sequences, and its one-way error is 1411 cm against our 17 cm.

VariantReal one-way, mean error (cm)Real round-trip, mean error (cm)Sim, mean ATE (cm)
Elevator-VIGS171529
Fold at the arrival877 (4 failed)378 (3 failed)284 (1 failed)
Fold at the arrival + w/o update projection (Elevator-LIO's design)1411 (3 failed)382 (4 failed)335 (1 failed)

The table covers the paper's 8 one-way and 8 round-trip real-world sequences and the 12 sim sequences. On a one-way sequence, the robot rides to another floor, and the error is the difference between the estimated height change and the laser-measured floor height. On a round-trip sequence, the robot rides away and returns to the starting floor, and the error is the estimated height difference on returning to the start. In simulation, the error is the ATE against the ground-truth trajectory. "(n failed)" gives the number of failed sequences, on which the estimate diverges or the run does not finish. Failed sequences are left out of the mean.

All three rows use our detector, our rests and our integration of the rise. They differ only in two design choices that we take from Elevator-LIO. The second row writes the rise into the poses at the arrival, as Elevator-LIO does. The third row combines both choices. It writes the rise into the poses at the arrival and lets the BA update the transport state, the state that holds the elevator's rise and vertical velocity. It is the direct adaptation of Elevator-LIO.

Benefit of the update projection. The update projection holds the transport state fixed through every optimization. The camera does not observe the transport state, so the update projection treats it as a consider parameter. Without the update projection, the BA shortens the online rise, the rise estimated during the ride, on all 40 rides.

Benefit of the deferred fold. During the ride, the rise is kept apart from the poses. The transport state holds it, and the poses stay at the height of the departure floor, because they describe the robot relative to the elevator. The fold is the step that adds the rise to the poses, so that the trajectory and the map move to the height of the arrival floor. Folding at the arrival adds the rise to keyframes that are still in the sliding window. The camera saw no rise between these keyframes, so their visual constraints say that they lie at the same height, and the next optimization removes most of the added rise again. Our deferred fold therefore waits until the keyframes from the elevator have left the sliding window. No visual constraint then contradicts the rise, so the whole rise stays in the trajectory. This lowers the one-way error from 877 to 17 cm and removes all 7 failures on the real-world sequences.

AuthorsOpenReview ↗

Response to Reviewer 7SUi (2/2): scope and the role of 3DGS

W2. Specific to elevators. We agree that the scope is elevators, as the Limitations state. Within this scope, the two assumptions hold for every elevator, and our method remains robust to elevator motion within lift standards and to the error levels of MEMS IMUs.

The two assumptions. Both are properties of the elevator, whose guide rails constrain it to vertical motion and whose rides begin and end at rest. We assume nothing else about how the elevator moves between the two rests. Its acceleration, its top speed, the duration of the ride and the number of floors can all vary, and the method needs no model of them.

Degradation in the sensitivity analysis. The degradation in App. D.8 comes from perturbations at or beyond the largest size that a standard elevator or a MEMS IMU reaches, and from moving the rests. Each perturbation is an IMU error or a motion of the elevator or the building, and we write it into the recorded IMU readings. Its size is its magnitude, for example the tilt angle or the noise density in the Size column below. We re-ran the study with every perturbation at this largest size. Each size follows the European lift standard EN 81-20:2020, the ride quality that lift makers specify, the comfort criteria for tall buildings, or the error levels of a consumer and an industrial MEMS IMU (Bosch BMI270, Analog Devices ADIS16448). The appendix cites the source of each size.

PerturbationSizeSim rides, median (cm)Sim rides, max (cm)Real rides, median (cm)Real rides, max (cm)
None (spread among unperturbed re-runs)–0.20.40.53.3
Horizontal oscillation of the elevator0.2 m/s² peak to peak0.10.3sim onlysim only
Horizontal oscillation of the building0.25 m/s² at 0.2 Hz4.310sim onlysim only
Tilt of the elevator0.15°1.22.6sim onlysim only
IMU rotated against the camera0.1°0.71.22.65.0
0.5°3.27.16.328
Gyroscope bias step at mid-ride4×10⁻³ °/s0.21.21.12.7
Accelerometer white noise2×10⁻³ m/s²/√Hz21524.616
Accelerometer bias random walk10⁻⁴ m/s³/√Hz8.2480.416
Accelerometer bias step at mid-ride2×10⁻³ m/s²2828.41429
Accelerometer bias drift3×10⁻⁵ m/s³9.2803.110
Accelerometer scale error0.5%18257.932
Vertical motion in the arrival rest5 mm at 3 Hz0.40.5sim onlysim only
20 mm2026sim onlysim only

Each cell is the rise change in cm, the absolute change of the estimated rise under the perturbation. The median and the maximum are taken over three sim rides of 33 to 50 m and four real-world rides of 14 to 50 m. "sim only" marks a perturbation that is applied only in simulation.

In summary, the rides in this study rise 14 to 50 m, and no perturbation changes this rise by more than 28 cm at the median. The motion of the elevator and of the building changes the rise by at most 10 cm. The IMU errors change it the most, by at most 14 cm at the median on the real-world rides and 28 cm on the sim rides.

The rests also need not be identified exactly. Our code now determines the rise from two rest windows, the few seconds at rest before the departure and after the arrival (see our general response). Moving either rest window 0.25 or 0.5 s earlier or later changes the rise by at most 2.3 cm at the median on the real-world rides and 12 cm on the sim ones.

More varied elevators and operating conditions. All our sequences are held out for the prompts and thresholds of the ride detector, which we developed on private recordings in three other buildings. One configuration serves 11 buildings of four types and reports no false ride on the public benchmarks. Three new real-world sequences add intermediate stops, a camera that starts moving before the elevator stops, and passengers who walk in front of the camera. All their rides are detected, and their one-way vertical errors stay within 2.1% of the floor-to-floor height.

Q2. Role of 3DGS. 3DGS renders the map and estimates neither the elevator's motion nor the robot's pose.

Role in tracking. Tracking and the rise come from the visual-inertial BA with the transport state, the state that holds the elevator's rise and vertical velocity. 3DGS takes no part in them.

Separating the two contributions. The transport state decides where the map is placed, and 3DGS decides how it looks. With the transport state, the two floors are mapped at their correct heights. Without it, VIGS-SLAM misses 14.9 of the 15.07 m rise on Campus2 (Fig. 6), so its map merges the two floors and paints the arrival floor's red doors onto a departure-floor wall. Its PSNR is 4.4 dB lower than ours (22.4 against 26.8 dB).

I appreciate the broad evaluation, including real and simulated elevator rides, elevator-free benchmarks, and ablation and sensitivity studies.

However, my main concern remains limited methodological novelty. The core approach adapts Elevator-LIO’s motion decomposition to monocular visual–inertial SLAM, and the new modules primarily address the engineering challenges of that adaptation. I also remain concerned about the role of 3DGS. The visual–inertial estimator alone separates elevator and robot motion. 3DGS is used for downstream mapping, and its potential contribution to motion separation remains unexplored. I therefore maintain my overall rating of 2.

Reviewer_yhWdOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper extends VIGS-SLAM so that a monocular camera + IMU system keeps tracking and mapping through elevator rides. In a moving cabin the camera only sees the robot’s motion relative to the cabin, while the IMU also senses the cabin’s motion, and existing estimators either stay on the departure floor or diverge. The authors express the robot pose in the cabin frame and add a per-keyframe transport state (rise h and vertical velocity u along gravity). This state is integrated from the vertical acceleration; its constant error is removed by zero-velocity constraints at the departure and arrival rests (Eq. 7), it is held fixed during BA (Eq. 8), and it is written into the poses once the ride keyframes leave the window. Rides are detected with SigLIP2 zero-shot prompts and an Omnidata depth-spread test, followed by IMU thresholds, and loop closure is blocked across floors. The evaluation uses 16 real handheld sequences (the rise is checked against laser-measured floor heights on the 8 one-way ones), 16 Isaac Sim sequences, and the four public benchmarks used by VIGS-SLAM. The method beats 13 baselines on the elevator sequences and stays close to VIGS-SLAM on the others.

Critical strengths
  • It documents a real failure case well. Table 8 and App. D.3 show that VIO, LIO and LVIO systems either stay pinned to the departure floor or diverge at the departure or arrival, and the proposed method completes all 28 elevator sequences. If released, the sequences would be useful for others working on multi-floor SLAM.

  • The analysis is careful. The observability argument (App. C.3) and the error model in Eq. 14, which matches the median effect of the bias and scale perturbations in Table 27 within 3 cm, make the limits of the method explicit. This is why my concerns are about the contribution and the evaluation rather than about correctness.

Critical weaknesses
  • Limited novelty: The BA never updates the transport state (Sec. 3.2.3), so the rise comes purely from double-integrating vertical acceleration between two zero-velocity rests. Pedestrian navigation already does this, and Elevator-LIO already proposed an elevator state with rest updates and a fold. What remains is mostly bookkeeping that keeps vision from undoing the IMU estimate.

  • Missing key baselines: VIGS-SLAM with the same detector and the same ZUPT-bounded rise added outside BA is the natural ablation, and without it the benefit of the BA formulation is unclear. A barometer-aided baseline is also missing. Elevator-LIO is never compared under the same detection.

  • Thin real-world evidence: The rise is measured on only 8 one-way sequences, and the onboard MindSLAM is more accurate on 3 of them. In simulation, the IMU is clean and the rides are idealized S-curves with clear rests, which favors this method.

  • Hand-tuned detector with no generalization test: Ride detection relies on 27 hand-written prompts and more than a dozen thresholds, and there is no held-out building. Table 24 shows a narrow margin: an elevator-probability threshold of 0.55 already misses the simulated glass rides.

Questions for authors
  • Q1: Results for the decoupled baseline in W2(a), and for Elevator-LIO when it receives both the departure and the arrival from your detector.

  • Q2: How are the prompts and thresholds selected, and on which sequences? Are there results on buildings that were not used for this?

  • Q3: In the glass elevators part of the image sees the building move, while the visual residual is written in the cabin frame and has zero derivative with respect to (h, u). How are these pixels handled? Station1, a glass elevator, has the largest relative error (78 cm on 7.29 m). Could the outside view be used to observe the rise?

  • Q4: Sec. 4.4 says tracking reduces to VIGS-SLAM when no ride is detected, but on EuRoC the ATE goes from 2.79 to 2.38 cm (V2_03: 3.27 to 1.39). The VIGS-SLAM numbers are copied from its paper (App. D.4), and Table 21 lists 4 armed windows on EuRoC, which change the keyframe rule (App. C.4). Can you re-run VIGS-SLAM in your setup and explain the difference?

2 comments
AuthorsOpenReview ↗

Response to Reviewer yhWd (1/2): novelty and the requested baselines

W1. Novelty. We agree that the rise is dead reckoning between two zero-velocity rests. The problem we solve is that this rise has to live inside a visual-inertial BA whose camera sees no rise. We answer the three parts of W1 in turn.

Pedestrian navigation. Pedestrian navigation integrates it from the IMU alone, with no camera in the estimator. In our BA, the camera saw no rise between the keyframes from the elevator, so their visual constraints say that they lie at the same height, and an optimization removes a rise that is added to these poses.

Elevator-LIO. We share Elevator-LIO's decomposition, but our input differs. Elevator-LIO needs a LiDAR, whereas we use only a monocular camera and an IMU. To our knowledge, no prior SLAM system handles elevator rides with these two sensors. With this input, Elevator-LIO's elevator state, rest updates and fold do not transfer as they are (next point). Beyond the sensors, we differ from it in the estimator and the ride detector.

  • Estimator. Elevator-LIO runs an iterated ESKF that keeps only the current state, whereas our dense BA optimizes a sliding window that still holds the ride keyframes after the arrival.
  • Ride detector. Our ride detector takes its visual semantics from frozen pretrained models and needs no training on elevator data. It detects all 24 of our real-world rides, against 15 for Elevator-LIO's LiDAR detector.

Elevator-LIO is also an unpublished arXiv preprint, which the ICLR 2027 Reviewer Guide does not require submissions to compare against.

"Mostly bookkeeping". Keeping vision from undoing the IMU estimate decides whether the method works. To show this, we adapted Elevator-LIO's design directly to our BA. This variant adds the rise to the poses at the arrival, as Elevator-LIO's fold does, and it lets the BA update the transport state, the state that holds the elevator's rise and vertical velocity. The next optimization then removes a median of 98% of the added rise. The variant fails on 7 of the paper's 16 real-world sequences, and on the sequences it completes, its one-way error is 1411 cm against our 17 cm.

W2, Q1. Requested baselines. We ran both requested baselines. On the real-world sequences, the decoupled baseline has much larger errors than Elevator-VIGS, and Elevator-LIO fails on 11 of 19 sequences.

MethodReal one-way, mean error (cm)Real round-trip, mean error (cm)Sim, mean ATE (cm)
Elevator-VIGS251529
Decoupled baseline, W2(a)403 (1 failed)187 (1 failed)27
Elevator-LIO, our departure and arrival, W2(c)24 (7 failed)93 (4 failed)644 (2 failed)

The table covers 11 one-way and 8 round-trip real-world sequences and 12 sim sequences. On a one-way sequence, the robot rides to another floor, and the error is the difference between the estimated height change and the laser-measured floor height. On a round-trip sequence, the robot rides away and returns to the starting floor, and the error is the estimated height difference on returning to the start. In simulation, the error is the ATE against the ground-truth trajectory. "(n failed)" gives the number of failed sequences, on which the estimate diverges or the run does not finish. Failed sequences are left out of the mean.

(a) Decoupled baseline. The decoupled baseline of W2(a) is VIGS-SLAM with our detector. Its BA has no transport state, and our rise is added to its poses outside the BA. During the ride, its inertial residuals still see the elevator's acceleration while its visual residuals see no rise, so the two compete inside the BA. In simulation, the visual residuals have high confidence and outweigh the inertial residuals, so they hold its poses in the elevator frame, and the decoupled baseline equals ours (27 against 29 cm). On 8 of the 19 real-world sequences, the inertial residuals dominate and its poses drift vertically. On most of them, the poses follow the elevator, so adding the rise counts it a second time. In glass elevators, for example, the visual confidence is as low as 0.04. In Elevator-VIGS, the transport state absorbs the elevator's acceleration, so the two residuals do not compete.

(b) Barometer. Our sequences have no barometer, and no compared system uses one. Because the elevator's motion is a separate state, any sensor that observes it, such as a barometer, can be added as a measurement of the rise hh without changing the visual residuals.

(c) Elevator-LIO under the same detection. Elevator-LIO fails on 11 of 19 real-world sequences. Although its elevator mode starts and ends at our departure and arrival, its filter diverges during the ride or while the elevator waits at an intermediate floor in 9 of 11.

AuthorsOpenReview ↗

Response to Reviewer yhWd (2/2): real-world evidence, detector, glass elevators, and EuRoC

W3. Real-world evidence. With our refined rise estimate and three new sequences, our error is lower than MindSLAM's on 10 of 11 one-way sequences, and the simulation is not idealized in the IMU or in the rests.

Only 8 one-way sequences, MindSLAM more accurate on 3. We refined our code so that it determines the rise from two rest windows, the few seconds in which the elevator stands still before it departs and after it arrives (see our general response). With this refinement, MindSLAM is more accurate than our method on only one of the paper's eight one-way sequences, Station1. We further recorded three new one-way sequences, and our error is lower than MindSLAM's on all three. Over the resulting 11 one-way sequences, our error is lower on 10 and averages 25 cm against MindSLAM's 185 cm, as in the results table of our general response.

Simulation. We answer the three points on the simulation.

  • IMU. As App. A.2 describes, all our sim IMU readings carry white noise at the densities measured on the Odin 1, a constant bias and a bias random walk, and the elevator vibrates vertically at 0.06 m/s² RMS.
  • S-curves. The elevator follows a jerk-limited S-curve, the profile that elevator drives commonly use, but our method assumes nothing about how the elevator moves between the two rests. Its acceleration, its top speed and the duration of the ride can all vary.
  • Rests. At every rest, the handheld robot keeps moving inside the elevator instead of standing still. Within each rest, it moves up and down by 1 to 4 cm at up to 10 cm/s, and its vertical velocity is about twice that in our real-world rides.

The rise is also stable under perturbations. In our re-run of App. D.8 up to the largest values in lift standards and IMU datasheets, the elevator's oscillation and tilt change the rise by at most 2.6 cm, the building's sway by at most 10 cm, and the IMU errors by at most 14 cm (real-world) and 28 cm (sim) at the median.

W4, Q2. Detector. All our benchmark sequences are held out for the prompts and thresholds of the ride detector, so every detection result in the paper is on a building that was not used to select them.

How and on which sequences (Q2). We chose them on seven unpublished recordings in three other buildings that vary the elevator's appearance and the camera's motion. The criterion was that the armed window, the interval in which the detector takes the robot to be inside an elevator, opens before every departure and never outside an elevator.

Buildings not used for the selection (Q2). None of the buildings in our benchmark was used. With one configuration, every ride in every sequence of the benchmark is detected, and no false ride is reported.

Margin in Table 24. The misses at 0.55 in Table 24 came from a wrongly rendered mesh in the sim glass elevator. With it fixed, every sim ride is detected up to 0.55 and every real-world ride, the five glass rides included, up to 0.8 (our threshold is 0.5).

Q3. Glass elevators. The BA gives little weight to the pixels that see the building move, so they do not disturb the rise. The large error on Station1 did not come from the glass.

How these pixels are handled. No dedicated term handles them. The BA weighs each visual residual by its confidence, and the pixels seen through the glass, which move differently from the elevator frame, lose on average 84% of their confidence once the elevator departs. The pixels on the elevator carry on average 6 times as much confidence. Vision therefore tracks the robot in the elevator frame, and the rise comes from the IMU.

Station1. The paper's 78 cm on Station1 comes from the departure our solve placed late on its slow start, not from the glass, and the rest windows fix this. Our code now determines the rise from the few seconds at rest before the departure and after the arrival (see our general response), so the exact instant of the departure no longer matters. On the glass elevators of Office2 and Station1, our rise errors are now −1.7 and −7.0 cm.

Could the outside view be used to observe the rise? We do not use the outside view yet and plan to use it in future work. The pixels seen through the glass show the building and could track the robot in the world frame, where their visual residuals depend on the rise. The pixels on the elevator track the robot in the elevator frame. Using both needs a way to tell which pixels lie outside the elevator.

Q4. EuRoC. The VIGS-SLAM paper and code differ. App. D.4 copied 2.79 cm from the paper, but the latest code is more accurate and agrees with ours.

  • Re-run. With the same TensorRT engines and three re-runs each, the latest VIGS-SLAM code gives 2.38 ± 0.02 cm and ours 2.41 ± 0.01 cm.
  • V2_03. Both reach 1.38 cm, against 3.27 cm in that paper.
  • Armed windows. The remaining difference comes from the armed windows. An armed window changes only the keyframe rule, and armed windows add keyframes only on MH_01 and MH_02 (ATE +0.36 and −0.13 cm).

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response to all reviewers: refined rise estimate, new sequences, and current results

We thank the three reviewers for their careful reviews. This general response summarizes what we refined in the method, which sequences we added, and the current main results. All our new results here and in our replies to each reviewer come from our latest code, which we will release.

Refinement: rest windows. The rise is integrated from the IMU between the departure (the instant the elevator starts to move) and the arrival (the instant it stops). The paper anchors this integration at these two instants. Our code now anchors it over two rest windows, the few seconds in which the elevator stands still before the departure and after the arrival, so the rise no longer depends on the exact instants.

New sequences. We recorded three new real-world sequences for the reviewers' questions. They add intermediate stops, a camera that starts moving before the elevator stops, and passengers who stand and walk in front of the camera. Our released dataset will include the new sequences.

Current results. The table gives the results on all our sequences, the three new real-world ones included, with the paper's metrics.

MethodReal one-way, mean error (cm)Real round-trip, mean error (cm)Sim, mean ATE (cm)
Elevator-VIGS251529
VIGS-SLAM2020 (1 failed)468908
MindSLAM18535–
Elevator-LIO44 (7 failed)78 (5 failed)222 (6 failed)
Elevator-LIO, our departure and arrival24 (7 failed)93 (4 failed)644 (2 failed)

The table covers 11 one-way and 8 round-trip real-world sequences and 12 sim sequences. On a one-way sequence, the robot rides to another floor, and the error is the difference between the estimated height change and the laser-measured floor height. On a round-trip sequence, the robot rides away and returns to the starting floor, and the error is the estimated height difference on returning to the start. In simulation, the error is the ATE against the ground-truth trajectory. "(n failed)" gives the number of failed sequences, on which the estimate diverges or the run does not finish. Failed sequences are left out of the mean.

The final version will include these results and clarifications.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.