Noctif3R: Feed-Forward Monocular Real-Time SLAM for Photon-Limited Scenes on Embedded Hardware

Mihir Chauhan; Aditya Uday Abhang; Kevin Biju Mathew; Aniket Bera

cs.ROSep 17, 2026

Noctif3R: Feed-Forward Monocular Real-Time SLAM for Photon-Limited Scenes on Embedded Hardware

Authors: Mihir Chauhan, Aditya Uday Abhang, Kevin Biju Mathew, Aniket Bera

Organizations: IDEAS Lab, Department of Computer Science, Purdue University, West Lafayette, IN, USA

Abstract

Robots carrying out tasks in dark environments need to localize from a single RGB camera, in light so low that the per-pixel signal approaches the sensor's own noise, on a power-constrained onboard computer, in real time. Each of these constraints has matured pipelines, but the intersection does not. Offline low-light reconstruction now recovers structure below -4 dB but is far too slow to run in real time, while the real-time monocular systems a robot can actually carry (DROID-SLAM, DPV-SLAM, etc.) degrade or fail when SNR gets low. We measured how they fail: across the nine lowest darkness levels of our scenes, DROID-SLAM returns a full-length trajectory carrying no information about the camera's motion on all nine, VGGT-SLAM and CUT3R on eight, pi^3 on seven, and DPV-SLAM on four. We present SYS, a monocular pipeline built on a low-light feed-forward pointmap front end with an explicit match gate, which returns three tracked trajectories and no uninformative ones, at the lowest error of any method where it tracks (24-47% of the no-information ceiling against 56-73% for the strongest baseline), and at the narrowest coverage. On a real robot video take in which 86.5% of delivered frames are entirely black, every configuration of ours stops after the lit beginning, while DROID-SLAM and DPV-SLAM each emit a pose for all 1178 frames. Our method contribution is an embedded execution path for the Jetson AGX Orin: running the map, keyframes and backend at 384 pixels with tracking at 256, together with two fixes to the per-frame pose solve, is a replicated Pareto improvement, 1.28x throughput at 0.964x error on one scene and 1.42x at 0.68x on a second, with 47% less peak GPU memory and 29% less energy per pose. We evaluate on a calibrated, bit-exact regenerable noise ladder, on relabelled real-world dark exposures, and on a new dark-room video ladder recorded from a Boston Dynamics Spot robot.

Figures & tables

Fig. 1: Noctif3R at a glance. (a) monocular RGB at the photon floor, no depth, stereo or IMU, shown with a display gain so the structure is visible; (b) the geometry and camera path recovered from those frames alone, on the bookshelf $-5$ dB rung of the calibrated ladder. Every candidate pair is gated on the match fraction $\phi$ of Eq. 1 : pairs that clear the gate become keyframes for a bounded back end, and pairs that do not produce no pose at all . That gate is why the system stops loudly instead of returning a confident trajectory with nothing in it (Sec. V-C ), and why its coverage is the narrowest of the methods we compare (Sec. VI-B ).

	previous	Noctif3R	ratio
poses/s	$2.512$	$\mathbf{3.204}$	$\mathbf{1.28\times}$
ATE (% of ceiling)	$21.47$	$\mathbf{20.70}$	$0.964\times$
peak CUDA (GiB)	$20.70$	$\mathbf{10.95}$	$-47\%$
energy per pose (J)	$14.84$	$\mathbf{10.47}$	$-29\%$
latency p95 (s)	$1.384$	$\mathbf{0.907}$	$-34\%$

TABLE I: The deployable configuration against the previous deployment, chapel

-10

dB ( synthetic , tripod stills), Jetson AGX Orin, MAXN . Three runs per arm; the two distributions are disjoint on both the throughput and the accuracy axis, with three-run ranges under

3\%

. ATE is the all-frame trajectory as % of the no-information ceiling.

rung	mean luma	frac $<5$	med. $\|\nabla I\|$	levels
0 (lights on)	$55.39$	$8.3\%$	$0.690$	$169$
1	$5.99$	$58.0\%$	$0.049$	$83$
2	$1.69$	$95.6\%$	$0.020$	$69$
3	$0.68$	$98.8\%$	$0.005$	$8$
4 (null)	$0.23$	$99.2\%$	$\mathbf{0.000}$	$\mathbf{1}$

TABLE II: Our Dark-room video ladder. Median

|\nabla I|

is the spatial gradient a matcher consumes. Rung 0 is a null-input control:

86.5\%

of its delivered frames are identically zero. These are H.264 with no calibrated SNR.

Panel of Figure 4 — Fig. 2: The Spot dark-room video ladder scene as delivered by the iPhone 15 Pro carried on the robot, with no display gain applied. (a) the lights-on control take (rung 0 of Table II ); (b) the same warehouse on a dark rung, where only the ceiling fixtures remain above the codec’s black level.

method	tracked	silent	loud
EC3R-SLAM [ 11 ]	5	4	0
DPV-SLAM [ 7 ]	4	4	1
Noctif3R , full	3	0	6
$\pi^{3}$ [ 9 ]	2	7	0
CUT3R [ 10 ]	1	8	0
VGGT-SLAM [ 8 ]	1	8	0

TABLE III: Tracked / silent / loud over the nine synthetic floor rungs (

-10/-15/-20

\times

three scenes); every-8 stills protocol.

method	poses	cov.	closure	outcome
Noctif3R , $k{=}1$	$20$	$0.017$	$0.972$	stops after preamble
Noctif3R , full	$20$	$0.017$	$0.905$	stops after preamble
Noctif3R , deploy	$22$	$0.019$	$0.999$	stops; $1149$ reloc fails
DROID-SLAM	$\mathbf{1178}$	$1.000$	$\mathbf{0.040}$	silent
DPV-SLAM	$\mathbf{1178}$	$1.000$	$0.354$	silent
VGGT-SLAM	$537$	$0.456$	n/a	crash: degenerate rot.

TABLE IV: The null-input control,

86.5\%

of whose delivered frames are identically black: the

1178

frames of the stride-two matrix, every method on identical input. Closure is return-to-start error over path length. No calibrated SNR.

configuration	poses/s	p50/p95/p99 (s)	W	J/pose	ATE
every-8th, colour $512$	$0.411$ – $0.413$	$0.416$ / $0.518$ / $1.54$	$34.8$ – $35.2$	$87$	$22.75\%$
every-frame, split $256$ / $512$	$2.54$	$0.166$ / $1.36$ / $1.85$	$36.2$	$14.8$	$21.47\%$
every-frame, floor mode	$2.485$	$0.312$ / $1.17$ / $1.74$	$39.6$	$16.1$	$40.1\%$
Noctif3R (map $384$ + solver fixes)	$\mathbf{3.204}$	$\mathbf{0.163}$ / $0.907$ / $1.17$	$\mathbf{32.6}$	$\mathbf{10.5}$	$\mathbf{20.70\%}$

TABLE V: Jetson AGX Orin, MAXN , chapel

-10

dB ( synthetic , tripod stills). Latency is acquisition-to-pose with a free-running source. Power is the whole-board INA3221 sum, mean over the run; ATE is the all-frame trajectory as % of the no-information ceiling. The first three rows are single runs; the last is the configuration of Sec. III-C , three-run mean.

lever	coverage	tracked	silent
as shipped	$16.5\%$	$1$	$0$
relocalization gate $0.30\!\to\!0.05$	$16.5\%$	$1$	$0$
tracking gate $0.05\!\to\!0.01$	$17.5\%$	$1$	$0$
bounded map re-initialization	$41$ – $62.5\%$	$1$ – $2$	$0$

TABLE VI: Config-only escape routes from the absorbing state, over six development floor rungs ( synthetic , every-eighth stills), each pre-registered before its runs and fitted on development scenes only. Coverage is pooled over the six rungs. No lever clears the pre-registered endpoint , which required two additional tracked rungs with no new silent failures.

Explore similar work

Jul 20, 2026

SLAM in Low-Light Environments: Project Report

Simultaneous localization and mapping (SLAM) is one of the fundamental problems in robotics, as it enables autonomous operations in real-world scenarios. Under low illumination, reduced contrast, sensor noise, and motion blur degrade both feature extraction and feature matching, while compensating with LiDAR, depth, or thermal sensors raises cost, power draw, and integration complexity. Existing benchmarks remain dominated by well-lit indoor or daylight sequences, leaving open how far SLAM with standard RGB cameras can be pushed in the dark. We benchmark six systems spanning the feature-based, direct, filter-based, and learning-based paradigms - ORB-SLAM3, DSO, Kimera-VIO, OpenVINS, DPVO, and DPV-SLAM - on five LaMARia sequences of varying difficulty and illumination, reporting absolute and relative pose error alongside control-point recall. Kimera-VIO is the only system to track all five sequences to completion, combining the lowest relative pose error with steadily growing absolute error due to the absence of loop closure; DPVO and DPV-SLAM never lose tracking but incur absolute errors of roughly 100 m under low light; and the classical monocular pipelines (ORB-SLAM3, DSO) together with the filter-based OpenVINS fail outright or diverge on most of the harder and low-light sequences. The results suggest that RGB-only SLAM maintains stable low-light tracking only when both inertial fusion and global optimization are present. Closing the remaining gap will likely require low-light-specific learned front-ends or a return to complementary sensing.

Oleh Basystyi, Anna Stasyshyn, Oleksandr Kosovan +1

Applied Sciences Faculty, Ukrainian Catholic University

Simultaneous Localization and Mapping Visual SLAM Visual-Inertial SLAM

Ask PDF

Mar 5, 2026

Dark3R: Learning Structure from Motion in the Dark

We introduce Dark3R, a framework for structure from motion in the dark that operates directly on raw images with signal-to-noise ratios (SNRs) below

-4

dB -- a regime where conventional feature- and learning-based methods break down. Our key insight is to adapt large-scale 3D foundation models to extreme low-light conditions through a teacher--student distillation process, enabling robust feature matching and camera pose estimation in low light. Dark3R requires no 3D supervision; it is trained solely on noisy--clean raw image pairs, which can be either captured directly or synthesized using a simple Poisson--Gaussian noise model applied to well-exposed raw images. To train and evaluate our approach, we introduce a new, exposure-bracketed dataset that includes

\sim

42,000 multi-view raw images with ground-truth 3D annotations, and we demonstrate that Dark3R achieves state-of-the-art structure from motion in the low-SNR regime. Further, we demonstrate state-of-the-art novel view synthesis in the dark using Dark3R's predicted poses and a coarse-to-fine radiance field optimization procedure.

Andrew Y Guo, Anagh Malik, SaiKiran Tedla +7

University of Toronto · Vector Institute · York University +3

Structure from Motion Camera Pose Estimation

Ask PDF

Sep 17, 2026

AMB3R-SLAM: Kilometer-scale SLAM with Hierarchical Backend

We present AMB3R-SLAM, a real-time monocular SLAM system capable of reconstructing kilometer-scale trajectories over 10k frames on a single consumer-grade GPU. Our model couples a lightweight front-end for low-latency online tracking with a hierarchical backend that progressively enforces local, mid-level, and global consistency. By avoiding bundle adjustment that relies on the static world assumption, our system naturally handles complex dynamic scenes out of the box. Furthermore, we demonstrate that our method can be extended to leverage stereo, RGB-D, and LiDAR as additional inputs. AMB3R-SLAM achieves strong camera tracking performance across 9 datasets, reducing the absolute trajectory error (ATE) of previous state-of-the-art methods on VBR and Oxford Spires by over 70%. With additional LiDAR input, our model further reduces ATE to sub-meter level on KITTI and VBR datasets.

Hengyi Wang, Lourdes Agapito

Department of Computer Science, University College London

Visual SLAM

Ask PDF