SHIFT: Surface-aware High-speed Integration For TSDFs

Ayaan Choudhury; Lokender Tiwari

cs.ROSep 6, 2026

SHIFT: Surface-aware High-speed Integration For TSDFs

Authors: Ayaan Choudhury, Lokender Tiwari

Organizations: Indian Institute of Technology Jodhpur · BotLab Dynamics

Abstract

Real-time 3D mapping is fundamental for autonomous robotic navigation, with Euclidean Signed Distance Fields (ESDFs) serving as the standard representation for online motion planning. While recent advancements in non- projective distance fields yield highly accurate maps, their computational overhead remains a severe bottleneck. Conventional integrators redundantly re-fuse millions of depth pixels every frame, even long after the corresponding voxels have converged, wasting significant computational resources in environments dominated by large planar surfaces. In this paper, we present SHIFT (Surface-aware High-speed Integration For TSDFs), an efficient mapping framework designed to reduce this per-frame update cost. By exploiting structural redundancy directly from 3D depth geometry, SHIFT compresses flat local regions into weighted super-rays and freezes flat-voxel gradients. A compact ESDF voxel layout further reduces the memory footprint of the remaining wavefront. Extensive evaluations across various RGB-D and LiDAR sequences show that SHIFT cuts TSDF cost by 1.42 to 4.07 times, while holding mesh error within millimeters, and reduces ESDF-layer memory by up to 28%

Figures & tables

Dataset	Sensor	Type	Frames	GT pts	Voxel (cm)
Flat [ 19 ]	RGB-D	synth.	399	3.20 M	5–25
C&L [ 13 ]	RGB-D	real	2701	0.64 M	5–25
TUM [ 20 ]	RGB-D	real	84	0.66 M	5–25
N. College [ 18 ]	LiDAR	real	1986	2.42 M	25–40

Table 1: Evaluation datasets. “Frames” counts integrated depth frames / scans. C&L is Cow & Lady.

Figure 2: Integration cost vs. voxel size, all methods. Top: TSDF time per depth frame. Bottom: ESDF time per ESDF refresh (not per depth frame). SHIFT (blue) is the lowest TSDF curve on every dataset and every resolution, and its ESDF cost tracks Voxfield while staying well below the projective baselines.

Figure 3: Accuracy vs. voxel size. SHIFT tracks Voxfield on every sequence and stays below the projective baselines on TSDF RMSE at 10 cm (RGB-D) and 25 cm (LiDAR).

Figure 4: Per-vertex Mesh $\rightarrow$ GT distance, Ours vs. Voxfield, on the Flat floor @ 10 cm and the N. College ground plane @ 25 cm.

		Chamfer-L1		Coverage		ESDF-GT
Dataset	Vox.	Ours	Vf.	Ours	Vf.	Ours	Vf.
Flat	10	3.82	3.71	87.7	88.0	7.23	6.72
Cow & Lady	10	7.44	7.54	82.1	81.7	9.54	9.17
TUM	10	4.74	4.71	99.7	99.7	8.24	8.37
N. College	25	10.37	10.39	94.7	94.6	23.35	24.08

Table 2: Chamfer-L1, reconstruction coverage, and ESDF-GT RMSE at the representative voxel (“Vox.”, cm). Chamfer-L1 and ESDF-GT in cm, coverage in %.

		TSDF		Backend		End-to-end
Dataset	Vox.	Vf.	Ours	Vf.	Ours	Vf.	Ours
Flat	5	26.70	9.99 (2.67 $\times$ )	29.95	13.87 (2.16 $\times$ )	41.56	25.25 (1.65 $\times$ )
	10	15.32	4.67 (3.28 $\times$ )	15.71	5.11 (3.08 $\times$ )	27.38	16.55 (1.66 $\times$ )
	15	13.38	3.62 (3.70 $\times$ )	13.53	3.79 (3.57 $\times$ )	25.28	15.13 (1.67 $\times$ )
	20	12.67	3.22 (3.94 $\times$ )	12.74	3.30 (3.86 $\times$ )	24.49	14.61 (1.68 $\times$ )
	25	12.36	3.04 (4.07 $\times$ )	12.40	3.09 (4.01 $\times$ )	24.07	14.51 (1.66 $\times$ )

Table 3: Per-frame runtime in ms for Voxfield (“Vf.”) and SHIFT (“Ours”) across voxel sizes (“Vox.”, cm), with SHIFT’s speedup in parentheses (

>\!1

means SHIFT is faster). “TSDF” is the integration stage the method modifies; “Backend” excludes the shared front end and end-to-end includes it.

Dataset	Vox.	ESDF layer (MB)			Process RSS (MB)
Dataset	Vox.	Ours	Voxfield	Ratio	Ours	Voxfield
Flat	10	109.7	118.8	0.92	278.0	237.6
Cow & Lady	10	118.8	148.5	0.80	373.4	373.1
TUM	10	119.1	148.8	0.80	283.8	280.4
N. College	25	1844.3	2551.5	0.72	2227.4	2840.9

Table 4: Memory footprint at the representative voxel size (“Vox.”, cm). “Ratio” is our ESDF layer over Voxfield’s.

Flat @ 10 cm (RGB-D)
Configuration	Total	TSDF	TSDF	Mesh
	(s)	(s)	RMSE	$\rightarrow$ GT
Voxfield	10.99	6.11	3.31	1.75
$+$ Classify only	10.87	6.30	3.48	1.90
$+$ Super-rays (SHIFT)	6.54	1.89	3.87	1.94
$-$ Skipping	6.55	1.89	3.88	1.94

Table 5: Component ablation on Flat and N. College. Times in s, errors in cm. Speedups are not repeated here; see Tab. 3 . Rows marked

+

add a stage cumulatively;

-

Skipping removes the skip test from full SHIFT and is not a further stage. Best among the three cumulative configurations in bold.

		Flat @ 10 cm		N. College @ 25 cm
Param.	Value	Total	RMSE	Total	RMSE
$\delta_{\mathrm{skip}}$ (cm)	0	6.55	3.88	104.4	12.82
	1	6.54	3.87	102.7	12.82
	3	6.54	3.82	104.1	12.83
Cell $P$	8	6.76	3.86	102.7	12.82
	12	6.54	3.87	110.3	12.98

Table 6: Hyperparameter sensitivity. Total time in seconds, TSDF RMSE in cm. Paper defaults are bold in the value column;

P{=}12

is the RGB-D default and

P{=}8

the LiDAR default.

Symbol / hyperparameter	RGB-D	LiDAR
Patch size $P$	12	8
Depth noise factor $\beta$	0.002	0.0015
Noise floor $\sigma_{\min}$ (m)	0.004	0.004
Edge depth jump $\tau_{0}$ (m)	0.04	0.08
Max. edge cell fraction $\varepsilon_{\mathrm{edge}}$	0	0.5
Connectivity angle $\theta_{\mathrm{lidar}}$ ( ∘ )	–	10

Table 7: Every classifier/fusion symbol from Section 3, mapped to its launch-time value.

Dataset	flat	edge	curved
Flat	78.9%	4.4%	16.7%
Cow & Lady	42.1%	2.6%	55.3%
TUM	47.3%	2.0%	50.7%
N. College	70.6%	8.8%	20.6%

Table 8: Per-pixel classifier label distribution, averaged over the logged frames of the representative-voxel run. LiDAR classifies more edge pixels than any RGB-D sequence.

Dataset	Method	Mean	Std	Min	Max
Flat	Voxfield	15.32	0.45	11.01	28.79
Flat	SHIFT	4.67	0.32	1.14	8.42
Cow & Lady	Voxfield	13.53	3.39	5.35	19.32
Cow & Lady	SHIFT	8.85	4.19	2.41	17.19
TUM	Voxfield	11.69	1.23	8.50	15.20
TUM	SHIFT	5.99	1.62	3.21	11.89

Table 9: Per-frame TSDF update time (ms) at the representative voxel size (10 cm), mean

\pm

std with [min, max] shown as separate columns.

Panel of Figure 5 — Figure 5: Stage breakdown across the ablation ladder for Flat and N. College. Without bundling, the classifier inflates the front end; bundling then recovers it in the TSDF stage.

Figure 6: Sensitivity of the three main hyperparameters (Table 6 in the main paper), normalized to the default configuration (

=100\%

). Red is TSDF time, blue is TSDF RMSE. Skip tolerance and freeze threshold are visually flat; LiDAR classifier cell size is the only knob with a clear effect.

Panel of Figure 7 — Figure 7: Front-end time (top) and Chamfer-L1 (bottom) vs. voxel size.

Panel of Figure 8 — Figure 8: TSDF coverage (top) and reconstruction coverage (bottom) vs. voxel size.

Panel of Figure 9 — Figure 9: ESDF ground-truth RMSE (top) and ESDF-layer memory (bottom) vs. voxel size. SHIFT is lower than Voxfield at every voxel size on every dataset for memory.

Panel of Figure 10 — Figure 10: Process RSS vs. voxel size.

3D Mapping

See all 32 papers

Low-Latency Mapless ESDF Queries for MAV Reactive Collision Avoidance
Oct 9, 2026Leonard Freißmuth, Cedric Le Gentil, Helen Oleynikova +1Collision Avoidance Robot Collision Avoidance
Mobile Robotics Lab, ETH Zurich
Tiered-Resolution Volumetric Mapping for Real-Time Simultaneous Planning and Mapping on Natural Outdoor Environments
Oct 14, 2026Afonso E. Carvalho, David Portugal, Paulo Peixoto +13D Mapping Robot Navigation
Institute of Systems and Robotics, University of Coimbra, Coimbra, Portugal · Robotics Institute, Carnegie Mellon University, Pittsburgh, PA, USA
D-BDM: A Direct and Efficient Boundary-Based Occupancy Grid Mapping Framework for LiDARs
Apr 14, 2026Benxu Tang, Yixi Cai, Fanze Kong +2Occupancy Grid Mapping Simultaneous Localization and Mapping
Mechatronics and Robotic Systems (MaRS) Laboratory, Department of Mechanical Engineering, The University of Hong Kong · Department of Robotics, Perception, and Learning, KTH Royal Institute of Technology

Oct 9, 2026J/K move · Enter open · S save

Low-Latency Mapless ESDF Queries for MAV Reactive Collision Avoidance

Leonard Freißmuth, Cedric Le Gentil, Helen Oleynikova, Stefan Leutenegger

Mobile Robotics Lab, ETH Zurich

Collision Avoidance Robot Collision Avoidance Signed Distance Functions

Open paper AskPDF arXiv

Reactive collision avoidance in unknown, cluttered environments must perceive obstacles and correct motion within milliseconds. Existing methods rely on maps for high-rate collision queries, but their acceleration structures require pre- processing. That cost is amortized in long-horizon planning, where a map serves millions of queries; reactive planning, however, is a near-field, short-horizon problem with potentially only hundreds of queries per measurement, where building a map may not pay off. We propose a mapless, Graphics Processing Unit (GPU)-accelerated Euclidean Signed Distance Field (ESDF) query with negligible pre-processing that operates directly on posed depth images, with resolution set by pooling the image rather than by a voxel size. On an embedded GPU it is simultaneously more accurate and lower in total latency than every voxel-grid configuration we measured between 2 cm and 20 cm, at one to two orders of magnitude less memory at matched resolution. With our query integrated into a Model Predictive Controller (MPC), simulated runs in laser-scanned forests at up to 9 m/s collide in 0.4 % through dense branches, against 5.9 % with the best voxel grid and 5.5 % with the best mapless baseline at our resolution, whose failures we attribute to voxel resolution and total latency respectively. We deploy the system onboard at two forest sites and fly at commanded speeds of up to 4 m/s.

Abstract

3D Mapping

Explore similar work

Low-Latency Mapless ESDF Queries for MAV Reactive Collision Avoidance

Tiered-Resolution Volumetric Mapping for Real-Time Simultaneous Planning and Mapping on Natural Outdoor Environments

D-BDM: A Direct and Efficient Boundary-Based Occupancy Grid Mapping Framework for LiDARs