Mamba-BEV: BEV Perception & SSM Fusion

• Mamba-BEV is a family of architectures that combine BEV representations with state-space models to efficiently handle spatial and temporal reasoning in autonomous driving.
• They employ novel serialization and adaptive scanning techniques to reconcile 2D BEV grids with 1D sequence models, reducing computational complexity compared to transformers.
• Empirical results demonstrate that tailored BEV-SSM fusion strategies boost accuracy in occupancy prediction, 3D detection, and end-to-end autonomous driving while enhancing efficiency.

Searching arXiv for papers on Mamba-BEV and related BEV state-space methods. arXiv search query: "Mamba BEV autonomous driving state space model" Mamba-BEV denotes a class of Bird’s Eye View perception architectures that combine BEV representations with Mamba or related state-space models (SSMs) for spatial or spatio-temporal reasoning. In the narrowest usage, it is the perception-and-temporal feature extraction backbone introduced inside ME$^3$-BEV, where surround-view RGB images are encoded into a unified BEV representation and then processed by a Mamba-based temporal model before reinforcement-learning control (Lu et al., 8 Aug 2025). In the broader literature, the same design pattern appears in occupancy prediction, camera-based 3D detection, multimodal fusion, online map construction, and end-to-end autonomous driving, where quadratic transformer attention or heavy dense 3D processing is replaced, approximated, or complemented by linear-time SSM-based operators adapted to BEV geometry (Tian et al., 2024, You et al., 2024, Ke et al., 18 Mar 2025, Yang et al., 27 Jul 2025, Wang et al., 8 Aug 2025, Hu et al., 12 Mar 2025).

1. Conceptual basis

The central motivation of Mamba-BEV systems is that BEV is a natural intermediate scene representation for driving, while standard transformer attention is expensive on large BEV token grids and over temporal windows. The literature therefore combines two ideas: a top-down BEV canvas for geometric consistency, and a sequence model with linear computational complexity for long-range context propagation. In ME$^3$-BEV, the BEV choice is justified by a unified coordinate system for roads, vehicles, lanes, obstacles, and navigation, along with improved translation equivariance and permutation invariance for downstream control (Lu et al., 8 Aug 2025). MambaOcc makes the same efficiency argument in occupancy prediction, explicitly replacing Transformer-style quadratic attention in the BEV encoder with linear Mamba-style attention (Tian et al., 2024).

Across these systems, the underlying dynamical form is an input-conditioned SSM. MambaOcc writes

$$h_t = \overline{\boldsymbol{A}} h_{t-1}+\overline{\boldsymbol{B}} x_t,\qquad y_t = \boldsymbol{C} h_t,$$

with

$$\bar{A}=\exp (\Delta A), \qquad \bar{B}=(\Delta A)^{-1}(\exp (\Delta A)-I)\cdot \Delta B,$$

and input-dependent parameters

$$B=s_{B}(x),\qquad C=s_{C}(x),\qquad \Delta = \tau_{\Delta}(W_{\Delta}+s_{\Delta}(x)).$$

ME$^3$-BEV formulates temporal fusion per modality as

$$\mathbf{h}_{m,k}^t=\mathbf{A}_k^t \mathbf{h}_{m,k}^{t-1}+\mathbf{B}_k^t \mathbf{h}_{k}^t,\qquad \mathbf{u}_k^t=\mathbf{C}_k^t \mathbf{h}_{m,k}^t+\mathbf{v}_t,$$

with input-dependent parameterization through $f_A$, $f_B$, and $f_C$ (Tian et al., 2024, Lu et al., 8 Aug 2025). The shared premise is that selective, input-dependent state updates can preserve long-range dependencies without the $^3$0 interaction pattern of self-attention.

2. Architectural pattern and representative systems

Most Mamba-BEV systems follow a recognizable pipeline: multi-view image or multimodal feature extraction, transformation into a BEV grid, optional temporal alignment or memory fusion, BEV-domain SSM processing, and a task-specific head. MambaOcc uses image feature extraction, LSS-style view transformation to BEV, a hybrid CNN–Mamba BEV encoder, and an occupancy head that recovers height through a channel-to-height rearrangement (Tian et al., 2024). MambaBEV uses image encoding, backward projection through spatial cross attention, a TemporalMamba module for history fusion, and a Mamba-based DETR head (You et al., 2024). MamBEV preserves a BEVFormer-like encoder-decoder structure but replaces core attention with Spatial Cross Mamba and Hydra/Mamba BEV self-attention (Ke et al., 18 Mar 2025).

Mamba-BEV also appears in multimodal and task-specialized systems. GMF-Drive fuses three forward-facing camera images and LiDAR through channel attention, a BEV-SSM block, and hierarchical deformable cross-attention (Wang et al., 8 Aug 2025). MambaMap uses a memory bank, BEV Mamba Fusion, and Instance Mamba Fusion for online vectorized HD map construction (Yang et al., 27 Jul 2025). DDHFusion places BEV-domain Mamba inside a broader dual-domain BEV-and-voxel detector, using cross-modal Mamba blocks for both homogeneous voxel fusion and homogeneous BEV fusion (Hu et al., 12 Mar 2025).

System	Primary task	Distinctive BEV/SSM mechanism
MambaOcc (Tian et al., 2024)	Occupancy prediction	VMamba-Tiny backbone, LSS BEV transform, hybrid CNN–Mamba encoder, LAR, channel-to-height head
MambaBEV (You et al., 2024)	Camera-based 3D detection	TemporalMamba with four-direction discrete rearrangement and Mamba-DETR
MamBEV (Ke et al., 18 Mar 2025)	Multi-camera 3D detection	Spatial Cross Mamba/XQSSM and Hydra BEV self-attention
MambaMap (Yang et al., 27 Jul 2025)	Online vectorized HD map construction	Memory bank with BEV Mamba Fusion and Instance Mamba Fusion
GMF-Drive (Wang et al., 8 Aug 2025)	End-to-end autonomous driving	BEV-SSM with polar positional encoding, raster/zigzag scans, adaptive fusion, HCA
DDHFusion (Hu et al., 12 Mar 2025)	Camera-LiDAR 3D detection	Cross-modal Mamba in both BEV and voxel domains
ME$^3$1-BEV (Lu et al., 8 Aug 2025)	End-to-end DRL driving	Lift-Splat BEV encoder with Mamba-based Temporal-Aware Fusion Module

3. Sequence order, geometry, and BEV-specific adaptations

A defining technical issue in Mamba-BEV research is that SSMs are sequence models, whereas BEV features are 2D spatial grids. The literature repeatedly identifies sequence-order sensitivity as a primary obstacle. MambaOcc makes this explicit: if a BEV grid is flattened into a sequence with an arbitrary order, spatially adjacent cells may be far apart in scan order, degrading recurrent propagation. Its Local Adaptive Reordering mechanism defines a remapping

$^3$2

with a relaxed many-to-one pseudo-reordering implemented by deformable convolution, so that Mamba receives a more spatially meaningful sequence (Tian et al., 2024).

MambaBEV addresses the same problem through a four-direction discrete rearrangement mechanism, serializing the BEV feature map in forward left, forward upward, reverse left, and reverse upward directions, then restoring and averaging the results. The paper states that it does not use snake-like serialization because of imbalance in adjacency interaction (You et al., 2024). MambaMap generalizes multi-directional scanning to left, right, up, and down on the BEV feature map, and additionally applies spatial-temporal scanning to instance-query sequences after Hungarian alignment of historical and current queries (Yang et al., 27 Jul 2025).

MamBEV treats the problem as one of geometry-aware cross-attention emulation. Rather than letting each BEV query interact with the entire image sequence, it projects BEV cells into camera views, keeps only in-bounds reference points, converts them to a 1D insertion order, and interleaves query tokens with image tokens in sequence space. Query tokens are assigned zero discretization update, so image tokens write to the SSM state and query tokens read from it (Ke et al., 18 Mar 2025). GMF-Drive adds a different class of BEV prior: ego-centric polar positional encoding with dimensional interleaving, two scan orders—raster and zigzag—and directional transition matrices with highest forward weight, moderate lateral weight, and lowest backward weight, together with distance decay

$^3$3

to encode the anisotropy of driving scenes (Wang et al., 8 Aug 2025). Taken together, these methods show that Mamba-BEV is not merely “Mamba instead of Transformer”; it is a family of sequence-construction schemes that try to reconcile 2D BEV geometry with 1D state-space scanning.

4. Task-specific realizations

In occupancy prediction, MambaOcc retains a BEV-based pipeline and avoids full dense 3D processing until the final stage. Multi-view images are encoded, transformed to BEV through Lift-Splat-Shoot, optionally fused over time by ego-motion alignment and concatenation of historical BEV features, and finally decoded through a channel-to-height occupancy head that predicts semantic voxel classes. The paper explicitly positions MambaOcc as the first Mamba-based occupancy network integrated into a BEV-based occupancy pipeline (Tian et al., 2024).

In camera-based 3D detection, two distinct Mamba-BEV strategies are visible. MambaBEV emphasizes temporal BEV fusion: current and historical BEV maps are aligned, fused by a convolutional block, rearranged into four directional sequences, and processed by Mamba2 before a Mamba-DETR head refines 900 object queries (You et al., 2024). MamBEV instead emphasizes the BEV encoder: Spatial Cross Mamba acts as a cross-attention analogue for BEV-query/image-feature interaction, and Hydra-based BEV self-attention refines the resulting grid before a BEVFormer-like decoder predicts 3D boxes and classes (Ke et al., 18 Mar 2025).

In multimodal 3D detection, DDHFusion shows that BEV-domain Mamba need not be used in isolation. The framework argues that BEV-only fusion is efficient but loses 3D detail, while voxel-only fusion preserves geometry but is expensive and awkward for query generation. Its solution is dual-domain homogeneous fusion: image features are lifted into both BEV and sparse voxels, voxel-domain Mamba blocks model global intra- and cross-modal relations, and BEV-domain Mamba blocks propagate dense context and address misalignment before progressive query generation and progressive decoding (Hu et al., 12 Mar 2025).

In end-to-end autonomous driving, the term “Mamba-BEV” is used most literally in ME$^3$4-BEV. The observation at time $^3$5 includes six surround-view RGB images of size $^3$6, road features $^3$7, vehicle features $^3$8, and navigation features $^3$9. A Lift-Splat style Spatial-Semantic Aggregator produces a $$h_t = \overline{\boldsymbol{A}} h_{t-1}+\overline{\boldsymbol{B}} x_t,\qquad y_t = \boldsymbol{C} h_t,$$0 BEV representation, modality-specific temporal features are processed by Mamba, and the fused representation feeds PPO Actor and Critic networks in a POMDP setting (Lu et al., 8 Aug 2025). GMF-Drive represents a different end-to-end design: three forward-facing camera images and raw LiDAR are fused through GM-Fusion, and a truncated diffusion decoder generates 8 waypoints over 4 seconds at 2 Hz (Wang et al., 8 Aug 2025).

In online vectorized HD map construction, MambaMap uses BEV as a temporal map-building canvas rather than a detection or occupancy substrate. A memory bank stores historical refined BEV features and instance queries; BEV Mamba Fusion warps historical BEV maps into the current ego frame and processes multi-directional scan sequences with a gated DSS block; Instance Mamba Fusion aligns historical queries by Hungarian matching and performs temporal modeling over aligned query sequences before DETR-style classification and regression of vectorized map elements (Yang et al., 27 Jul 2025).

5. Empirical results and efficiency

Mamba-BEV methods are typically evaluated on the joint axes of accuracy and efficiency. On Occ3D-nuScenes, MambaOcc reports 43.4 mIoU for MambaOcc† with 893.8 GFLOPs and 79.5M parameters, compared with FlashOcc†:M3 at 43.3 mIoU, 1467.5 GFLOPs, and 137.1M parameters; MambaOcc-Large† reaches 44.1 mIoU with 1002 GFLOPs and 119M parameters. The paper states that, relative to FlashOcc, MambaOcc reduces parameters by 42% and computational costs by 39%, while MambaOcc-Large improves by $$h_t = \overline{\boldsymbol{A}} h_{t-1}+\overline{\boldsymbol{B}} x_t,\qquad y_t = \boldsymbol{C} h_t,$$1 mIoU with 14% fewer parameters and 32% fewer FLOPs (Tian et al., 2024). Its BEV-encoder ablation reports 34.16 mIoU for a pure CNN encoder, 34.72 for pure SS2D, 34.93 for a CNN-SS2D hybrid, and 35.41 for LAR-SS2D; temporal fusion raises results from 35.41 mIoU for MambaOcc to 39.78 mIoU for MambaOcc-4D (Tian et al., 2024).

On nuScenes 3D detection, MambaBEV-B with VoV-99 at $$h_t = \overline{\boldsymbol{A}} h_{t-1}+\overline{\boldsymbol{B}} x_t,\qquad y_t = \boldsymbol{C} h_t,$$2 reports NDS $$h_t = \overline{\boldsymbol{A}} h_{t-1}+\overline{\boldsymbol{B}} x_t,\qquad y_t = \boldsymbol{C} h_t,$$3 and mAP $$h_t = \overline{\boldsymbol{A}} h_{t-1}+\overline{\boldsymbol{B}} x_t,\qquad y_t = \boldsymbol{C} h_t,$$4, while the R-101 version reports NDS $$h_t = \overline{\boldsymbol{A}} h_{t-1}+\overline{\boldsymbol{B}} x_t,\qquad y_t = \boldsymbol{C} h_t,$$5, mAP $$h_t = \overline{\boldsymbol{A}} h_{t-1}+\overline{\boldsymbol{B}} x_t,\qquad y_t = \boldsymbol{C} h_t,$$6, and the lowest mAVE $$h_t = \overline{\boldsymbol{A}} h_{t-1}+\overline{\boldsymbol{B}} x_t,\qquad y_t = \boldsymbol{C} h_t,$$7 among the reported settings. The paper also reports that four-direction rearrangement improves over single-direction serialization, and that convolution-based temporal fusion is better than simple concatenation, particularly for velocity error (You et al., 2024). MamBEV-Small with 4 frames reports NDS $$h_t = \overline{\boldsymbol{A}} h_{t-1}+\overline{\boldsymbol{B}} x_t,\qquad y_t = \boldsymbol{C} h_t,$$8 and mAP $$h_t = \overline{\boldsymbol{A}} h_{t-1}+\overline{\boldsymbol{B}} x_t,\qquad y_t = \boldsymbol{C} h_t,$$9. Its efficiency analysis shows that at $$\bar{A}=\exp (\Delta A), \qquad \bar{B}=(\Delta A)^{-1}(\exp (\Delta A)-I)\cdot \Delta B,$$0 BEV and $$\bar{A}=\exp (\Delta A), \qquad \bar{B}=(\Delta A)^{-1}(\exp (\Delta A)-I)\cdot \Delta B,$$1 images, XQSSM requires 51 GFLOPs and 6.5 GB, deformable attention 49.5 GFLOPs and 4.9 GB, and dot-product attention 1432.5 GFLOPs and more than 24 GB; the appendix further reports 4.7 FPS and 2.2 GB for Spatial Cross Mamba versus 3.7 FPS and 9.8 GB for a Transformer (Ke et al., 18 Mar 2025).

DDHFusion reports 71.6 mAP and 73.8 NDS on nuScenes validation, with DDHFusion-large reaching 72.8 mAP and 74.4 NDS. Its ablations show gains from both voxel-domain and BEV-domain Mamba: adding homogeneous voxel fusion raises a 68.4 mAP / 71.2 NDS baseline to 70.0 / 72.1, adding homogeneous BEV fusion yields 69.1 / 71.8, and combining both reaches 70.6 / 72.6 before further gains from progressive query generation and progressive decoding (Hu et al., 12 Mar 2025).

For end-to-end autonomous driving, GMF-Drive reports PDMS 88.9 on NAVSIM navtest, compared with 88.1 for DiffusionDrive, with NC 98.2, DAC 97.3, TTC 94.2, Comfort 100, and EP 83.3. The same paper shows that a naive SSM replacement is insufficient: C-EffiMamba drops to 88.02 PDMS, whereas HCA+BEV-SSM reaches 88.69 (Wang et al., 8 Aug 2025). In CARLA, ME$$\bar{A}=\exp (\Delta A), \qquad \bar{B}=(\Delta A)^{-1}(\exp (\Delta A)-I)\cdot \Delta B,$$2-BEV reports, under low-density traffic, collision rate 0.26 versus 0.81 for e2e-CLA and driving score 71.2 versus 39.7; under high-density traffic, collision rate 0.43 versus 0.86 and driving score 55.2 versus 36.8. The combined system stabilizes around 400,000 steps, while SSA-only and TAFM-only variants converge around 800,000 steps (Lu et al., 8 Aug 2025).

For online vectorized HD map construction, MambaMap reports 67.3 mAP at $$\bar{A}=\exp (\Delta A), \qquad \bar{B}=(\Delta A)^{-1}(\exp (\Delta A)-I)\cdot \Delta B,$$3 and 66.5 mAP at $$\bar{A}=\exp (\Delta A), \qquad \bar{B}=(\Delta A)^{-1}(\exp (\Delta A)-I)\cdot \Delta B,$$4 on the original nuScenes validation set, 64.9 and 60.7 mAP on Argoverse2 at the same respective ranges, 40.1 mAP on a geographically disjoint nuScenes split, and 61.0 mAP on a disjoint Argoverse2 split. Its ablations show that adding BEV Mamba Fusion alone raises mAP from 61.5 to 65.2, adding Instance Mamba Fusion alone reaches 63.7, and combining both yields 67.3. Multi-directional BEV scanning is best at 67.3, and among state-space choices DSS performs best at 67.3, exceeding Mamba at 65.3 and Mamba-2 at 65.2 (Yang et al., 27 Jul 2025).

6. Limitations, misconceptions, and open directions

A recurring misconception is that Mamba-BEV is simply a drop-in replacement of transformer attention by a generic SSM. Several papers argue against that interpretation directly. GMF-Drive reports that a plain SSM replacement hurts performance relative to transformer-like fusion, and that gains appear only after introducing BEV-specialized priors such as directional sequencing, polar positional encoding, distance decay, gating, and adaptive fusion (Wang et al., 8 Aug 2025). MamBEV similarly argues that naive Mamba-style cross-attention would require storing all image tokens in the state and would recover transformer-like quadratic behavior; its contribution is a geometry-aware query insertion scheme that avoids that failure mode (Ke et al., 18 Mar 2025). MambaMap adds a related caution from the model-selection side: although the framework is conceptually inspired by Mamba, the actual best-performing block on its task is DSS rather than Mamba or Mamba-2 (Yang et al., 27 Jul 2025).

Another consistent limitation is sequence-order sensitivity. MambaOcc, MambaBEV, and MambaMap each devote dedicated mechanisms to scan ordering or adaptive reindexing, which indicates that BEV serialization remains a first-order design variable rather than an implementation detail (Tian et al., 2024, You et al., 2024, Yang et al., 27 Jul 2025). MamBEV’s reliance on camera calibration and geometry-derived reference points is an additional assumption in cross-modal image-to-BEV fusion (Ke et al., 18 Mar 2025). In ME$$\bar{A}=\exp (\Delta A), \qquad \bar{B}=(\Delta A)^{-1}(\exp (\Delta A)-I)\cdot \Delta B,$$5-BEV, evaluation is simulation-only in CARLA, and the paper notes fixed sensing with six surround RGB cameras and limited analysis of robustness to sensor failure or noise (Lu et al., 8 Aug 2025). MambaBEV also implies a cost-accuracy trade-off: more rearrangement directions and higher BEV resolution improve results but add parameters and computation (You et al., 2024).

These patterns suggest a stable interpretation of the field. Mamba-BEV is best understood not as a single architecture, but as a research program centered on BEV-native state-space modeling. The dominant questions are how to serialize BEV features, how to encode driving-specific spatial priors, how to fuse modalities without recovering attention-like complexity, and how to balance BEV efficiency against the loss of vertical or instance-level structure. The literature to date indicates that the strongest results arise when SSMs are coupled with explicit geometric priors, task-specific scan strategies, and carefully designed interfaces between BEV grids, history, and downstream prediction heads (Tian et al., 2024, Wang et al., 8 Aug 2025, Ke et al., 18 Mar 2025).