GRS-SLAM3R: Real-Time Dense SLAM with Gated Recurrent State

핵심 요약

GRS-SLAM3R은 DUSt3R식 pointmap prediction을 real-time RGB-only dense SLAM으로 확장하기 위해 latent memory를 gate로 관리하고 긴 trajectory를 submap 단위로 정렬한다.

문제pairwise 3D만으로는 SLAM 부족 해결gated state + submap 근거ATE / reconstruction / FPS

한 문장 요약

이 논문은 gated recurrent spatial memory와 keyframe 기반 local/inter-submap alignment를 결합해 연속 RGB frame에서 dense world-frame point cloud와 camera pose를 실시간으로 추정한다.

Contribution 01

Gated Recurrent State

reset/update gate로 어떤 memory token을 억제, 보존, 업데이트할지 결정.

Contribution 02

World-frame Dense Output

RGB에서 local pointmap, world-frame pointmap, confidence, 6-DoF pose를 함께 출력.

Contribution 03

Hierarchical Submap Alignment

submap별 memory reset과 pose-graph registration으로 recurrent drift 전파를 제한.

Contribution 04

Real-time Dense SLAM

single RTX 4090에서 online dense reconstruction과 pose estimation을 real-time 수준으로 수행.

핵심 인사이트

핵심은 submap 내부에서는 gated token update로, submap 사이에서는 명시적 alignment로 recurrent drift를 두 단계에서 제한하는 점이다.

처리 흐름

01RGB Streammonocular input

02Encoderframe tokens

03Reset Gatememory filtering

04Decodermemory-image fusion

05Update Gatestate update

06Submap Alignglobal consistency

접근 방식 비교

DUSt3R / MASt3R

pairwise 3D prior

pointmap prediction은 강하지만 online memory와 global SLAM consistency를 단독으로 보장하기 어려움.

CUT3R / SLAM3R

continuous reconstruction

online reconstruction으로 확장하지만 긴 sequence의 spatial memory와 drift control이 여전히 어려움.

GRS-SLAM3R

gated memory + submaps

선택적 recurrent memory update와 local/global submap alignment를 결합해 RGB-only dense SLAM으로 구성.

논문 상세 정리

이하에서는 배경지식, notation, 보조 유도까지 포함해 논문의 세부 내용을 살펴본다.

Problem: DUSt3R 계열을 왜 바로 SLAM으로 쓰기 어려운가

GRS-SLAM3R의 출발점은 DUSt3R 기반 end-to-end reconstruction이 camera calibration 없이 dense 3D를 잘 예측하지만, 대개 image pair 또는 short sequence 중심이라 online SLAM에 필요한 장기 spatial memory와 global consistency가 부족하다는 점이다.

Figure 1. Scene Reconstruction Performance.large-scale multi-room apartment scene에서 RGB-only 입력으로 reconstruction quality와 completeness를 유지하는지를 시각적으로 보여준다.

Problem Flow

문제는 “좋은 pointmap prior” 자체보다, 그 prior를 시간적으로 누적하고 전체 map으로 묶는 방식에 있다.

01Dense SLAM target

pose와 dense geometry를 RGB stream에서 동시에 얻어야 함.

02Pairwise limitation

image pair prediction은 spatial memory와 long-term consistency를 직접 제공하지 않음.

03Memory contamination

오래된 memory나 noisy observation이 누적되면 latent state가 drift함.

04Submap need

긴 sequence에서는 하나의 recurrent state만으로 global drift를 제한하기 어려움.

Related Work 맥락

관련 연구를 어떻게 나눠 볼까

논문은 dense monocular SLAM, 3D reconstruction, DUSt3R 기반 online reconstruction을 비교하며, 자신들의 위치를 real-time RGB-only dense SLAM으로 잡는다.

계열	강점	GRS-SLAM3R의 문제의식
Sparse / feature SLAM	빠르고 안정적인 tracking	dense geometry가 부족해 planning/collision avoidance에 한계.
NeRF / 3DGS SLAM	고품질 dense representation	depth, intrinsics, per-scene optimization에 의존하거나 실시간성이 약할 수 있음.
DUSt3R / MASt3R	camera-free pointmap prediction	pairwise 3D prior를 online SLAM memory로 확장해야 함.
CUT3R / SLAM3R	continuous reconstruction 방향	spatial memory update/forgetting과 long sequence drift를 더 직접적으로 다룰 필요.

Mechanism: gated recurrent state와 submap이 어떻게 맞물리나

방법론은 두 축으로 정리된다. Frontend는 gated recurrent state로 frame-level pointmap과 pose를 예측하고, backend는 keyframe/submap 구조로 local alignment와 inter-submap pose graph를 수행한다.

Figure 2. System overview.왼쪽은 gated recurrent state를 이용한 dense reconstruction frontend, 오른쪽은 keyframe/submap 기반 hierarchical alignment를 보여준다.

Mechanism Thread Summary

GRS-SLAM3R은 하나의 latent state를 무작정 누적하지 않고, gate로 memory를 선택적으로 업데이트한 뒤 submap 단위로 drift를 끊는다.

모듈	담당하는 문제	출력/효과
Encoder	현재 frame을 token representation으로 변환	$F_t$
Reset gate	현재 관측과 결합하기 전에 irrelevant memory 억제	$M_t^{\mathrm{reset}}$
Decoder	memory token, image token, pose token 사이 정보 교환	candidate memory와 enriched feature
Update gate	새 memory와 이전 memory의 비율 결정	$M_t$
Submap alignment	long sequence drift를 submap 단위로 제한	local map과 global pose graph

1. Frame token과 gate 계산

입력은 monocular RGB sequence $\{I_t\}_{t=1}^{N}$이고, 목표 출력은 dense point cloud $X_t$, confidence $C_t$, camera pose $P_t$다. 먼저 encoder가 현재 frame token을 만든다.

$$ F_t=\operatorname{Encoder}(I_t) \tag{1} $$

Eq. (1). Frame token encoding.현재 RGB image $I_t$를 latent memory와 상호작용할 token feature $F_t$로 변환한다.

$$ R_t=G_r(M_{t-1},F_t),\qquad U_t=G_u(M_{t-1},F_t) \tag{2} $$

Eq. (2). Recurrent gate weights.이전 memory와 현재 feature를 함께 사용해 reset gate $R_t$와 update gate $U_t$의 가중치를 계산한다.

Figure 3. Gate structure.reset/update gate는 memory token과 current frame token을 함께 보고, suppress할 과거 정보와 carry할 정보를 나눠 결정한다.

2. Memory reset, decoder, update

Reset gate는 오래되었거나 현재 관측과 맞지 않는 memory component를 줄인다. 그 뒤 decoder는 reset memory, pose token, frame token을 함께 처리해 candidate memory와 frame feature를 갱신한다.

$$ M_t^{\mathrm{reset}}=R_t\odot M_{t-1} \tag{3} $$

Eq. (3). Reset memory filtering.reset gate를 elementwise로 적용해 현재 관측과 맞지 않는 memory 성분을 억제한다.

$$ [\hat{M}_t,z'_t\oplus F'_t] =\operatorname{Decoder}([M_t^{\mathrm{reset}},z_t\oplus F_t]) \tag{4} $$

Eq. (4). Joint token decoding.reset된 memory token, pose token, image token을 decoder에서 함께 처리해 candidate memory와 갱신된 frame feature를 만든다.

$$ M_t=U_t\odot\hat{M}_t+(1-U_t)\odot M_{t-1} \tag{5} $$

Eq. (5). Gated memory update.update gate가 candidate memory와 이전 memory의 비율을 정해 장기 spatial memory를 선택적으로 갱신한다.

3. Pointmap / pose heads

Decoder 이후에는 local camera frame과 global world frame의 pointmap을 모두 예측한다. 논문이 DUSt3R와 달리 world coordinate pointmap을 직접 출력한다고 강조하는 이유가 여기에 있다.

$$ \hat{X}_t^{\mathrm{self}},\hat{C}_t^{\mathrm{self}} =\operatorname{Head}_{\mathrm{self}}(F'_t) \tag{6} $$

Eq. (6). Local pointmap head.갱신된 frame feature에서 local camera frame 기준의 dense pointmap과 confidence를 예측한다.

$$ \hat{X}_t^{\mathrm{world}},\hat{C}_t^{\mathrm{world}} =\operatorname{Head}_{\mathrm{world}}(F'_t,z'_t) \tag{7} $$

Eq. (7). World pointmap head.frame feature와 pose token을 사용해 world coordinate의 metric-scale pointmap과 confidence를 직접 예측한다.

$$ \hat{P}_t=\operatorname{Head}_{\mathrm{pose}}(z'_t) \tag{8} $$

Eq. (8). Pose head prediction.갱신된 pose token에서 현재 frame의 6-DoF camera pose를 예측한다.

4. Hierarchical submap alignment

새 frame이 마지막 keyframe이나 submap anchor와 충분히 겹치지 않으면 keyframe 또는 새 submap을 만든다. 이렇게 submap 단위로 latent memory를 reset하면 long-horizon recurrent drift가 전체 trajectory로 계속 전파되는 것을 줄일 수 있다.

$$ \mathcal{L}_{\mathrm{local}}^{S} =\sum_{e\in \mathcal{E}_{l}^{S}}\sum_{v\in e}\sum_{i=1}^{HW} C_i^{v,e} \left\lVert \xi_i^{v,S}-\sigma_e^S P_e^S X_i^{v,e} \right\rVert \tag{9} $$

Eq. (9). Intra-submap alignment loss.submap 내부의 pairwise pointmap을 local coordinate에 맞춰 latent state reset 이후의 local map을 정렬한다.

$$ \min_{\{T_s\}} \sum_{(u,v)\in \mathcal{E}} \left\lVert \operatorname{Log}(\Delta T_{u,v}^{-1}T_u^{-1}T_v) \right\rVert_{\Sigma}^{2} + \left\lVert \operatorname{Log}(T_{s_0}^{-1}\bar{T}_{s_0}) \right\rVert_{\Sigma_0}^{2} \tag{10} $$

Eq. (10). Inter-submap pose graph.submap pose를 node로 두고 adjacent 및 loop constraint를 edge로 삼아 전체 map의 global consistency를 맞춘다.

Training loss와 notation

학습 objective는 무엇을 맞추나

학습은 CUT3R을 따라 confidence-weighted 3D regression loss와 pose loss를 사용해 geometry와 camera motion을 함께 감독한다.

기호	의미	역할
$\hat{x}_i$, $x_i$	예측/GT 3D point	scale-normalized pointmap supervision.
$c_i$	confidence	hard pixel 영향 조절, confidence collapse는 $-\beta\log c_i$로 방지.
$\hat{q}_t$, $\hat{\tau}_t$	예측 pose의 rotation/translation	pose supervision을 rotation과 scale-normalized translation으로 분리.

$$ L_{\mathrm{regr}} =\sum_{i=1}^{M} \left( c_i\left\lVert \frac{\hat{x}_i}{\hat{s}}-\frac{x_i}{s}\right\rVert_2 -\beta\log c_i \right) \tag{11} $$

Eq. (11). Scale-aware point regression.confidence가 있는 3D point 예측을 GT와 scale-normalized 좌표에서 맞추고 confidence collapse를 억제한다.

$$ L_{\mathrm{pose}} =\sum_{t=1}^{N} \left( \left\lVert\hat{q}_t-q_t\right\rVert_2 + \left\lVert\frac{\hat{\tau}_t}{\hat{s}}-\frac{\tau_t}{s}\right\rVert_2 \right) \tag{12} $$

Eq. (12). Pose supervision loss.rotation quaternion과 scale-normalized translation을 함께 비교해 camera pose 예측을 감독한다.

Evidence: pose, reconstruction, runtime을 어떻게 검증했나

평가는 camera pose accuracy, surface reconstruction, real-time performance, ablation으로 구성된다. Test dataset은 NRGBD, 7-Scenes, Apartment, NES이며, tracking은 ATE-RMSE, reconstruction은 accuracy/completeness를 중심으로 본다.

Evaluation Roadmap

이 논문은 task를 dataset 이름보다 “SLAM 시스템으로서 무엇을 만족하는가”에 맞춰 읽는 편이 좋다.

Tracking

7-Scenes ATE-RMSE로 frame별 pose 안정성 확인.

Dense map

NRGBD, 7-Scenes, Apartment, NES에서 reconstruction quality 확인.

System evidence

FPS와 ablation으로 gate/local alignment/submap의 역할 확인.

Pose Evidence

Table I. Pose estimation on 7Scenes dataset.7-Scenes에서 per-scene ATE RMSE를 cm 단위로 보고하며, GRS-SLAM3R은 평균 8.27cm로 CUT3R과 MASt3R-SLAM보다 낮다.

Surface Reconstruction Evidence

Table II. Reconstruction results on the NRGBD dataset.NRGBD에서 mean accuracy/completion이 CUT3R보다 낮아 sparse sampling과 작은 overlap에서도 recurrent memory가 geometry를 유지함을 보인다.

Table III. Reconstruction results and runtime on 7Scenes. — Table III. Reconstruction results and runtime (FPS) on 7Scenes.GRS-SLAM3R은 7-Scenes에서 reconstruction accuracy/completion과 약 15 FPS의 runtime을 함께 보고해 accuracy-speed trade-off를 강조한다.

Figure 4. Qualitative scene reconstruction results on chess seq-05 in 7Scenes. — Figure 4. Qualitative scene reconstruction results.7Scenes chess seq-05에서 red outline은 낮은 predictive accuracy, green outline은 높은 accuracy, yellow는 ground truth를 나타내며, 제안 방법이 local geometry를 더 안정적으로 유지하는 예시다.

Figure 5. Qualitative scene reconstruction results on the NES dataset. — Figure 5. Qualitative scene reconstruction results.NES dataset의 긴 corridor와 큰 trajectory에서 기존 방법은 drift, breaking, scale shrinkage를 보이지만, GRS-SLAM3R은 submap alignment로 corridor structure와 scale을 더 안정적으로 유지한다.

Table IV. Quantitative results on the Apartment dataset.약 $100m^2$ 규모의 multi-room scene에서 accuracy, completion, ATE를 함께 보고한다. SLAM3R은 explicit camera pose를 출력하지 않아 ATE가 보고되지 않는다.

Runtime / Ablation Evidence

Figure 6. Ablation on Gated Recurrent model.gate가 없으면 latent state를 제대로 조절하지 못해 inconsistent geometry가 누적되고, gated update는 새 observation을 선택적으로 통합한다.

Table V. Ablation of Gate, Local Align, and Submap.Gate, local align, submap을 모두 제거하면 error와 ATE가 크게 악화되고, 세 구성요소를 함께 사용할 때 가장 좋은 결과를 보인다.

Usage / Limits: 어떤 상황에 잘 맞나

GRS-SLAM3R은 RGB-only 입력으로 dense geometry와 pose를 동시에 얻고 싶고, offline global alignment보다 빠른 online 동작이 필요한 상황에 잘 맞는다. 반면 하나의 실시간 시스템이므로, 매우 큰 loop closure나 domain gap이 큰 outdoor scene에서는 submap registration 품질과 covisibility threshold를 따로 확인해야 한다.

When to Use / Avoid

논문의 실험과 구조를 application 조건으로 정리하면 다음과 같다.

구분	요약	이유
Good fit	RGB-only dense indoor SLAM	depth나 camera intrinsics 없이 dense point cloud와 pose를 함께 출력.
Good fit	real-time reconstruction이 중요한 경우	offline global alignment 계열보다 빠른 online pipeline을 목표로 함.
Check carefully	long corridor, large loop, weak overlap	submap split과 pose graph constraint 품질이 전체 drift를 좌우.
Limitation	정밀한 metric map이 최우선인 경우	online speed와 dense reconstruction을 함께 노리므로 추가 refinement가 필요할 수 있음.

Problem: why is DUSt3R-style 3D not enough for SLAM?

GRS-SLAM3R starts from the gap between strong pairwise pointmap prediction and online dense SLAM. DUSt3R-style methods can recover dense 3D without camera calibration, but online SLAM also needs long-term spatial memory and global consistency.

Problem Flow

The hard part is not only predicting a pointmap, but accumulating it over time and organizing it as a global map.

01Dense SLAM target

Estimate pose and dense geometry from RGB streams.

02Pairwise limitation

Pairwise 3D prediction does not provide online memory by itself.

03Memory contamination

Noisy observations or outdated memory can drift the latent state.

04Submap need

A single recurrent state is not enough to control long-range drift.

Related Work context

How to place this paper

The paper compares dense monocular SLAM, 3D reconstruction, and DUSt3R-based online reconstruction, then positions itself as real-time RGB-only dense SLAM.

Lineage	Strength	GRS-SLAM3R's concern
Sparse / feature SLAM	fast and stable tracking	sparse maps are insufficient for dense geometry use cases.
NeRF / 3DGS SLAM	high-quality dense representation	often relies on depth, intrinsics, optimization, or limited speed.
DUSt3R / MASt3R	camera-free pointmap prediction	pairwise 3D priors must be extended into online SLAM memory.
CUT3R / SLAM3R	continuous reconstruction	memory update, forgetting, and long-sequence drift need stronger handling.

Mechanism: how do gated state and submaps work together?

The method has two main axes. The frontend predicts frame-level pointmaps and pose with a gated recurrent state. The backend uses keyframes and submaps for local alignment and inter-submap pose-graph registration.

Mechanism Thread Summary

GRS-SLAM3R avoids blindly accumulating one latent state: gates control memory updates, and submaps break long-horizon drift.

Module	Problem handled	Output / effect
Encoder	turns the current frame into tokens	$F_t$
Reset gate	suppresses irrelevant memory before fusion	$M_t^{\mathrm{reset}}$
Decoder	mixes memory, image, and pose tokens	candidate memory and enriched features
Update gate	balances new and previous memory	$M_t$
Submap alignment	limits long-sequence drift	local maps and global pose graph

1. Frame tokens and gates

The input is a monocular RGB sequence $\{I_t\}_{t=1}^{N}$. The outputs are dense point cloud $X_t$, confidence $C_t$, and camera pose $P_t$.

$$ F_t=\operatorname{Encoder}(I_t) \tag{1} $$

Eq. (1). Frame token encoding.The current RGB image is encoded into frame tokens that interact with the recurrent latent memory.

$$ R_t=G_r(M_{t-1},F_t),\qquad U_t=G_u(M_{t-1},F_t) \tag{2} $$

Eq. (2). Recurrent gate weights.Previous memory and current features jointly determine the reset and update gate weights.

2. Memory reset, decoder, update

The reset gate suppresses outdated memory, the decoder exchanges information between memory/image/pose tokens, and the update gate blends candidate memory with previous memory.

$$ M_t^{\mathrm{reset}}=R_t\odot M_{t-1} \tag{3} $$

Eq. (3). Reset memory filtering.The reset gate suppresses memory components that are outdated or inconsistent with the current observation.

$$ [\hat{M}_t,z'_t\oplus F'_t] =\operatorname{Decoder}([M_t^{\mathrm{reset}},z_t\oplus F_t]) \tag{4} $$

Eq. (4). Joint token decoding.The decoder processes reset memory, pose tokens, and frame tokens to produce candidate memory and updated features.

$$ M_t=U_t\odot\hat{M}_t+(1-U_t)\odot M_{t-1} \tag{5} $$

Eq. (5). Gated memory update.The update gate blends candidate and previous memory so long-range spatial memory changes selectively.

3. Pointmap / pose heads

After decoding, the network predicts both local-camera and world-frame pointmaps. This is why the method fits online SLAM better than pairwise pointmap-only reconstruction.

$$ \hat{X}_t^{\mathrm{self}},\hat{C}_t^{\mathrm{self}} =\operatorname{Head}_{\mathrm{self}}(F'_t) \tag{6} $$

Eq. (6). Local pointmap head.The head predicts a dense pointmap and confidence in the local camera frame.

$$ \hat{X}_t^{\mathrm{world}},\hat{C}_t^{\mathrm{world}} =\operatorname{Head}_{\mathrm{world}}(F'_t,z'_t) \tag{7} $$

Eq. (7). World pointmap head.Frame features and the pose token directly produce metric-scale world-frame pointmaps and confidence.

$$ \hat{P}_t=\operatorname{Head}_{\mathrm{pose}}(z'_t) \tag{8} $$

Eq. (8). Pose head prediction.The updated pose token is decoded into the current frame's 6-DoF camera pose.

4. Hierarchical submap alignment

Frames are promoted to keyframes and new submaps according to covisibility with the last keyframe and current submap anchor. Each submap keeps its own latent memory and local coordinate frame.

$$ \mathcal{L}_{\mathrm{local}}^{S} =\sum_{e\in \mathcal{E}_{l}^{S}}\sum_{v\in e}\sum_{i=1}^{HW} C_i^{v,e} \left\lVert \xi_i^{v,S}-\sigma_e^S P_e^S X_i^{v,e} \right\rVert \tag{9} $$

Eq. (9). Intra-submap alignment loss.Pairwise pointmaps are aligned inside each submap to refine the local map after state resets.

Eq. (10). Inter-submap pose graph.Submap poses are optimized with adjacent and loop constraints to preserve global consistency.

Training loss and notation

What is supervised?

The paper follows CUT3R with confidence-weighted 3D regression and pose loss. These are supporting details relative to the gated/submap mechanism.

Symbol	Meaning	Role
$\hat{x}_i$, $x_i$	predicted / GT 3D point	scale-normalized pointmap supervision.
$c_i$	confidence	down-weights hard points while $-\beta\log c_i$ avoids collapse.
$\hat{q}_t$, $\hat{\tau}_t$	rotation / translation of predicted pose	pose supervision with normalized translation.

$$ L_{\mathrm{regr}} =\sum_{i=1}^{M} \left( c_i\left\lVert \frac{\hat{x}_i}{\hat{s}}-\frac{x_i}{s}\right\rVert_2 -\beta\log c_i \right) \tag{11} $$

Eq. (11). Scale-aware point regression.Predicted and ground-truth 3D points are matched after scale normalization while confidence collapse is penalized.

$$ L_{\mathrm{pose}} =\sum_{t=1}^{N} \left( \left\lVert\hat{q}_t-q_t\right\rVert_2 + \left\lVert\frac{\hat{\tau}_t}{\hat{s}}-\frac{\tau_t}{s}\right\rVert_2 \right) \tag{12} $$

Eq. (12). Pose supervision loss.The loss supervises camera pose with quaternion rotation error and scale-normalized translation error.

Evidence: how are pose, reconstruction, and runtime evaluated?

The evaluation covers camera pose, surface reconstruction, real-time performance, and ablations. The main metrics are ATE-RMSE for tracking and accuracy/completeness for reconstruction.

Evaluation Roadmap

The results are easier to read by SLAM capability rather than by dataset name.

Tracking

ATE-RMSE on 7-Scenes.

Dense map

Reconstruction quality on NRGBD, 7-Scenes, Apartment, and NES.

System evidence

FPS and ablations for gate, local alignment, and submaps.

Pose Evidence

Surface Reconstruction Evidence

Runtime / Ablation Evidence

Usage / Limits: when is it useful?

GRS-SLAM3R is useful when RGB-only dense geometry and pose are needed online. For very large loops, weak-overlap scenes, or strong domain shift, submap registration quality and covisibility thresholds should be checked carefully.

When to Use / Avoid

The experiments and system design suggest the following application conditions.

Category	Summary	Reason
Good fit	RGB-only dense indoor SLAM	outputs dense point clouds and pose without depth or intrinsics.
Good fit	real-time reconstruction	targets an online pipeline faster than offline global alignment.
Check carefully	long corridors, large loops, weak overlap	submap split and pose-graph constraints control global drift.
Limitation	high-precision metric mapping as the only priority	online speed and dense reconstruction may still require refinement.

Wooseok Song

GRS-SLAM3R: Real-Time Dense SLAM with Gated Recurrent State

핵심 요약

Gated Recurrent State

World-frame Dense Output

Hierarchical Submap Alignment

Real-time Dense SLAM

pairwise 3D prior

continuous reconstruction

gated memory + submaps

논문 상세 정리

Problem: DUSt3R 계열을 왜 바로 SLAM으로 쓰기 어려운가

Related Work 맥락

Mechanism: gated recurrent state와 submap이 어떻게 맞물리나

Training loss와 notation

Evidence: pose, reconstruction, runtime을 어떻게 검증했나

Usage / Limits: 어떤 상황에 잘 맞나

Problem: why is DUSt3R-style 3D not enough for SLAM?

Related Work context

Mechanism: how do gated state and submaps work together?

Training loss and notation

Evidence: how are pose, reconstruction, and runtime evaluated?

Usage / Limits: when is it useful?

Comments

모듈	담당하는 문제	출력/효과
Encoder	현재 frame을 token representation으로 변환	\(F_t\)
Reset gate	현재 관측과 결합하기 전에 irrelevant memory 억제	\(M_t^{\mathrm{reset}}\)
Decoder	memory token, image token, pose token 사이 정보 교환	candidate memory와 enriched feature
Update gate	새 memory와 이전 memory의 비율 결정	\(M_t\)
Submap alignment	long sequence drift를 submap 단위로 제한	local map과 global pose graph

기호	의미	역할
\(\hat{x}_i\), \(x_i\)	예측/GT 3D point	scale-normalized pointmap supervision.
\(c_i\)	confidence	hard pixel 영향 조절, confidence collapse는 \(-\beta\log c_i\)로 방지.
\(\hat{q}_t\), \(\hat{\tau}_t\)	예측 pose의 rotation/translation	pose supervision을 rotation과 scale-normalized translation으로 분리.