A Survey of Spatial Memory Representations for Efficient Robot Navigation

Ma. Madecheen S. Pangaliman; Steven S. Sison; Erwin P. Quilloy; Rowel Atienza

cs.CVApr 13, 2026

A Survey of Spatial Memory Representations for Efficient Robot Navigation

Authors: Ma. Madecheen S. Pangaliman, Steven S. Sison, Erwin P. Quilloy, Rowel Atienza

Organizations: University of the Philippines Diliman · University of Santo Tomas

Abstract

As vision-based robots navigate larger environments, their spatial memory grows without bound, eventually exhausting computational resources, particularly on embedded platforms (8-16GB shared memory, $<$ 30W) where adding hardware is not an option. This survey examines the spatial memory efficiency problem across 88 references spanning 52 systems (1989-2025), from occupancy grids to neural implicit representations. We introduce the $α= M_{\text{peak}} / M_{\text{map}}$ , the ratio of peak runtime memory (the total RAM or GPU memory consumed during operation) to saved map size (the persistent checkpoint written to disk), exposing the gap between published map sizes and actual deployment cost. Independent profiling on an NVIDIA A100 GPU reveals that $α$ spans two orders of magnitude within neural methods alone, ranging from 2.3 (Point-SLAM) to 215 (NICE-SLAM, whose 47,MB map requires 10GB at runtime), showing that memory architecture, not paradigm label, determines deployment feasibility. We propose a standardized evaluation protocol comprising memory growth rate, query latency, memory-completeness curves, and throughput degradation, none of which current benchmarks capture. Through a Pareto frontier analysis with explicit benchmark separation, we show that no single paradigm dominates within its evaluation regime: 3DGS methods achieve the best absolute accuracy at 90-254,MB map size on Replica, while scene graphs provide semantic abstraction at predictable cost. We provide the first independently measured $α$ reference values and an $α$ -aware budgeting algorithm enabling practitioners to assess deployment feasibility on target hardware prior to implementation.

Figures & tables

Panel of Figure 1 — Figure 1: Evolution of spatial memory representations along two competing demands: geometric completeness and memory efficiency. Sparse features ( 1 ) traded completeness for efficiency; neural methods ( 2 ) recovered completeness via learned compression; scene graphs ( 3 ) added semantic abstraction. The scene graph box is dashed to indicate that the 48 MB figure reflects only the graph abstraction layer [ 26 ] ; the required metric-semantic backend [ 54 ] adds scene-dependent cost (dashed arrow), shifting the true system footprint leftward. The scene graph’s vertical position reflects its semantic completeness (objects, rooms, places) rather than geometric fidelity.

Panel of Figure 2 — Figure 2: Taxonomy of spatial memory representations with representative citations and typical efficiency metrics (gray). ATE = absolute trajectory error on standard benchmarks; map sizes are representative for building-scale environments. 3DGS = 3D Gaussian Splatting. Semantic maps include open-vocabulary approaches leveraging vision-language models such as CLIP.

System	ATE	Map	Peak	FPS	Eff.
	(cm)	(MB)	(MB)
VINS-Mono [ 52 ]	10.6 $\\|$	40 $*$	150 $*$	20 $*$	2.4
Basalt [ 69 ]	8.8 $\lx@sectionsign$	35 $*$	120 $*$	30 $*$	3.2
ORB-SLAM3 [ 5 ]	3.5 ${\dagger}$	55 ${\ddagger}$	220 ${\ddagger}$	30	5.2

Table 1: Feature-based visual SLAM comparison on EuRoC [ 4 ] (MH01–MH05; sensing modes vary per system; see footnotes). Eff. =

10^{3}/(\text{ATE}\times M_{\text{map}})

; see Section 4 for caveats on this composite metric.

System	ATE	PSNR	Map	Peak	Eff.
	(cm)	(dB)	(MB)	(MB)
iMAP [ 63 ]	3.12 f	22.1 d	1 a	— d	320.5
NICE-SLAM [ 86 ]	1.06 b	24.4 g	47 e	10,082 e	20.1
Co-SLAM [ 71 ]	1.00 h	30.2 h	8 e	1,258 e	125.0
Point-SLAM [ 56 ]	0.52	35.2	2,865 e	6,563 e	0.7
SplaTAM † [ 29 ]	0.36	34.1	254 e	14,024 e	10.9

Table 2: Neural SLAM on Replica (synthetic). Eff. =

10^{3}/(\text{ATE}\times M_{\text{map}})

, higher is better.

\dagger

: CVPR/ECCV 2024 or later.

System	Paradigm	ATE (cm)	PSNR (dB)	Map (MB)	Peak (MB)	FPS	$\alpha$
Sparse — evaluated on EuRoC [ 4 ] (real-world, stereo-inertial)
ORB-SLAM3 [ 5 ]	Sparse	3.5	—	55	220 ${\dagger}$	30	4.0 ${\dagger}$
VINS-Mono [ 52 ]	Sparse (VI)	10.6 ${\ddagger}$	—	40	—	20	—
Basalt [ 69 ]	Sparse (VI)	8.8 $\lx@sectionsign$	—	35 $\lx@sectionsign$	120 $\lx@sectionsign$	30 $\lx@sectionsign$	$\sim$ 3.4 $\lx@sectionsign$
Learned Flow — evaluated on EuRoC / TUM RGB-D (real-world)
DROID-SLAM [ 64 ]	Learned Flow	2.0 $\star$	—	— $\star\star$	11,000 $\star\star$	6	—

Table 3: Unified efficiency analysis across spatial memory paradigms. Overhead factor

\alpha=M_{\text{peak}}/M_{\text{map}}

;

\alpha_{\text{CPU}}

denotes CPU RSS,

\alpha_{\text{GPU}}

denotes GPU allocation (see Section 4 ). Cross-benchmark caveat: sparse systems are evaluated on EuRoC (real-world, stereo-inertial) and neural systems on Replica (synthetic, RGB-D); ATE and memory values are not directly comparable across horizontal dividers. Within-benchmark comparisons are valid; cross-paradigm conclusions should be drawn from

\alpha

and scaling behavior. Per-value provenance notes are in Tables 1 and 2 .

Panel of Figure 3 — Figure 3: Accuracy vs. map size ( $M_{\text{map}}$ ) by benchmark. Left: Sparse (EuRoC, linear scale). Right: Neural (Replica, log scale); dashed = Pareto front. Lower-left is better. Point-SLAM’s measured 2.9 GB checkpoint (vs. 80 MB survey estimate) moves it far right, off the Pareto front. Plotting $M_{\text{peak}}$ instead would further shift high- $\alpha$ systems rightward.

Panel of Figure 4 — Figure 4: Runtime GPU memory on Replica/room0 (1 Hz sampling, A100-SXM4-80GB, baseline subtracted). All four systems independently measured ( $\diamond$ ). Co-SLAM stays flat at ${\sim}1.3$ GB (bounded hash). NICE-SLAM oscillates between 3–7 GB due to periodic global mapping passes (absolute peak 10.1 GB, Table 3 ). Point-SLAM stabilizes at 5–6 GB after initialization. SplaTAM grows monotonically to ${\sim}12.7$ GB as Gaussians accumulate (peak 14.0 GB).

Constraint	Recommended	$\alpha$ range	Type	Avoid
CPU-only ( $<$ 8 GB)	Sparse, Octree	3–5 a	CPU	Neural
Embedded GPU ( $<$ 16 GB)	Sparse, SG	4–10 $*$	CPU	Raw 3DGS
Dense geometry	TSDF, 3DGS	${\sim}$ 55 b	GPU	Sparse only
Photo rendering	3DGS, NeRF	2–215 c	GPU	Occupancy
Multi-hour	Submaps, Stream	varies	—	Monolithic
Semantic	SG, VLM	4–10 $*$	CPU	Geom. only

Table 4: Representation selection guide by primary constraint.

System	Update	Forget	Part.	$\alpha$	$\Delta M$
ORB-SLAM3 [ 5 ]	Incr+BA	Culling	Sub	4	— a
RTAB-Map [ 35 ]	Incr	WM $\to$ LTM	Sub	—	Bound. b
DSO [ 14 ]	Window	Temporal	Mono	—	—
MS-SLAM [ 82 ]	Window	Sparsif.	Mono	—	$>$ 70% c
Co-SLAM [ 71 ]	Incr	None	Mono	157	—
GigaSLAM [ 10 ]	Incr	LOD	Hier	—	LOD d

Table 5: Memory dynamics taxonomy. Consol. = consolidation, Hier = hierarchical, Mono = monolithic.

\Delta M

= reported memory reduction.

\alpha

from Table 3 where available; note that forgetting reduces

M_{\text{map}}

but not

\alpha

Robotic Spatial Memory

See all 21 papers

Memory for Attention: Language-Conditioned Re-Perception with a Vision--Language--Motion Map
Jul 26, 2026Dibyendu GhoshActive Perception Visual Attention
Dibyendu Ghosh
Semantic Zone-Based Map Management for Stable AI-Integrated Mobile Robots
Mar 31, 2026Huichang Yun, Seungho YooMobile Robotics Simultaneous Localization and Mapping
Dept. of computer Engineering, Pukyong National University, Busan, Korea
Spatial Memory by Agents, for Agents
Oct 15, 2026Qianwei Wang, Ce Zhang, Zhanpeng Luo +3Robotic Spatial Memory Agent Memory
Qianwei Wang is with Princeton University · Ce Zhang is with Carnegie Mellon University · Zhanpeng Luo is with Johns Hopkins University +2

Jul 26, 2026J/K move · Enter open · S save

Memory for Attention: Language-Conditioned Re-Perception with a Vision--Language--Motion Map

Dibyendu Ghosh

Active Perception Visual Attention VLMs for Robotics Robotic Perception Memory-Augmented Agents

Open paper AskPDF arXiv

A robot carrying a persistent, behavior-annotated map faces two planning questions, and its memory answers only one well. The spatial-navigation question -- how to walk around a room -- we address first and report a negative: building on Vision--Language--Motion Maps (VLMM), a behavior-aware planner cost cuts a planning-time objective by

\sim

35% over 28 AI2-THOR scenes, but under closed-loop execution the real benefit nearly vanishes (

\sim

4%) and an on-demand vision--language model (VLM) does as well. The resource-allocation question differs: under a limited perception budget, what should the robot re-observe now to keep its map fresh? Framing re-perception as this attention decision, we show a persistent map's memory (change-history, or even just recency of last sighting) yields the best schedule (held-out), matching an oracle, while the memoryless VLM prior is poor. Because the schedule reallocates budget toward what matters, memory's benefit concentrates on the important objects (

\sim

1.6

\times

the mean), and a downstream fetch task confirms fewer wasted trips; the gain grows with per-instance heterogeneity exactly as a Cauchy--Schwarz bound predicts -- it equals

\mathrm{Var}(\sqrtλ)

, the variance of root-volatility. With a real CLIP prior on rendered objects the advantage is

+21

26\%

. The map's distinctive value appears when the task is language-conditioned: told what to track, VLMM grounds the relevant objects (open-vocabulary) and tracks their change (memory), beating even a strong relevance-weighted recency baseline (

+2.5\%

) -- so its motion channel adds value beyond a last-seen timestamp -- and an on-demand VLM (

+8.9\%

); neither language nor dynamics alone suffices. The map earns its keep not by telling the robot how to walk around a room, but by telling it what to pay attention to.

Abstract

Robotic Spatial Memory

Explore similar work

Memory for Attention: Language-Conditioned Re-Perception with a Vision--Language--Motion Map

Semantic Zone-Based Map Management for Stable AI-Integrated Mobile Robots

Spatial Memory by Agents, for Agents