CBDES MoE: Modular BEV Perception

• The paper introduces a hierarchically decoupled MoE framework that replaces a fixed image backbone with a heterogeneous expert pool, achieving a 1.6-point mAP and 4.1-point NDS boost on nuScenes.
• CBDES MoE is a modular system that decouples the perception stage to enable dynamic, input-aware routing via a Self-Attention Router for efficient expert specialization.
• The design preserves established BEV fusion and prediction pipelines while offering scalability for future extensions across other functional modules in autonomous driving.

Searching arXiv for the primary paper and closely related BEV/MoE context. arXiv search: "CBDES MoE Hierarchically Decoupled Mixture-of-Experts for Functional Modules in Autonomous Driving" Computing Brain DEvelopment System Mixture-of-Experts (CBDES MoE) is a hierarchically decoupled Mixture-of-Experts (MoE) architecture for functional modules in autonomous driving, introduced for Bird’s Eye View (BEV) perception systems based on multi-sensor feature fusion. It is designed to address limited input adaptability, constrained modeling capacity, and suboptimal generalization in existing multi-modal BEV methods by replacing a fixed image backbone with a heterogeneous expert pool and a lightweight Self-Attention Router (SAR) that performs dynamic expert path selection and sparse, input-aware inference. The framework is presented as the first modular MoE constructed at the functional module granularity within the autonomous driving domain, and its reported evaluation on nuScenes shows consistent gains over fixed single-expert baselines in 3D object detection (Xiang et al., 11 Aug 2025).

1. Conceptual definition and scope

CBDES MoE is organized around a functional decomposition of a BEV multi-sensor perception stack into three modules: perception, fusion, and prediction. In the reported instantiation, only the perception module is decoupled into experts, while the fusion and prediction modules remain single-path. More specifically, the camera branch replaces a single fixed image backbone with a heterogeneous pool of $K$ experts, each independently extracting $2$D feature maps from raw multi-view images; the fusion module projects each expert’s output into the BEV grid and fuses image, LiDAR, and radar BEV features; and the prediction module applies a standard detection head for classification and regression on the fused BEV representation (Xiang et al., 11 Aug 2025).

The term hierarchical decoupling refers to the possibility that each functional module may have its own expert set and routing logic. Within the current scope, however, the method applies this decoupling only at the perception stage. This distinction is central to interpreting the framework correctly: CBDES MoE is not presented as a fully expert-routed end-to-end pipeline across all modules, but as a module-level MoE architecture whose current implementation targets the image-backbone stage. A plausible implication is that the design aims to capture most of the representational benefit of modular specialization while preserving the stability and compatibility of a conventional BEV fusion and detection stack.

2. Hierarchically decoupled architecture at the perception stage

At the perception stage, the input image tensor or pre-normalized feature is denoted by

$$F_{\mathrm{img}} \in \mathbb{R}^{B \times 3 \times H \times W}.$$

Each expert $\mathcal{E}_k$ produces a feature map

$$F^{(k)} = \mathcal{E}_k(F_{\mathrm{img}}) \in \mathbb{R}^{B \times C \times H' \times W'}.$$

Adapter layers consisting of $1 \times 1$ convolution and batch normalization ensure that all expert outputs share the same $(C, H', W')$, after which the routed combination is formed as

$$F_{\mathrm{fused}} = \sum_{k=1}^{K} r_k F^{(k)}, \qquad r_k = \mathbf{P}_{b,k}.$$

This construction makes the MoE operate at the level of a functional subnetwork rather than at the level of individual tokens or layers. In the reported system, the expert pool is structurally heterogeneous and contains four different backbone families: Swin Transformer, ResNet, ConvNeXt, and Pyramid Vision Transformer (PVT). The paper characterizes these as, respectively, hierarchical windowed self-attention, classical residual CNN, modernized convolutional design, and PVT. The explicit use of heterogeneous experts, rather than multiple replicas of a single architecture, is one of the defining properties of CBDES MoE (Xiang et al., 11 Aug 2025).

A common misunderstanding is to interpret the method as merely ensembling backbones. The reported formulation differs from a fixed ensemble because the router generates image-level soft assignments over experts and can optionally collapse them to top-$1$ routing at inference. This suggests an input-conditional computation graph rather than a static average of multiple models.

3. Self-Attention Router (SAR) gating mechanism

SAR is the gating component that computes image-level soft assignments over the expert pool. Given an input feature map

$$\mathbf{X} \in \mathbb{R}^{B \times C \times H \times W},$$

the router proceeds through hierarchical convolution-pooling embedding, self-attention encoding, and expert scoring (Xiang et al., 11 Aug 2025).

The hierarchical embedding stage is defined as

$2$0

$2$1

$2$2

where each $2$3 is followed by BatchNorm and PReLU, $2$4 is $2$5 max pooling, and $2$6 is the token dimension, with $2$7 given as an example.

The resulting tensor is flattened into a token sequence,

$2$8

with $2$9. A single layer of multi-head self-attention is then applied using learnable projections $$F_{\mathrm{img}} \in \mathbb{R}^{B \times 3 \times H \times W}.$$0:

$$F_{\mathrm{img}} \in \mathbb{R}^{B \times 3 \times H \times W}.$$1

$$F_{\mathrm{img}} \in \mathbb{R}^{B \times 3 \times H \times W}.$$2

$$F_{\mathrm{img}} \in \mathbb{R}^{B \times 3 \times H \times W}.$$3

Global pooling over tokens yields

$$F_{\mathrm{img}} \in \mathbb{R}^{B \times 3 \times H \times W}.$$4

The scoring stage passes $$F_{\mathrm{img}} \in \mathbb{R}^{B \times 3 \times H \times W}.$$5 through a $$F_{\mathrm{img}} \in \mathbb{R}^{B \times 3 \times H \times W}.$$6-layer MLP with PReLU:

$$F_{\mathrm{img}} \in \mathbb{R}^{B \times 3 \times H \times W}.$$7

Here $$F_{\mathrm{img}} \in \mathbb{R}^{B \times 3 \times H \times W}.$$8 is the soft assignment weight of expert $$F_{\mathrm{img}} \in \mathbb{R}^{B \times 3 \times H \times W}.$$9 for input $\mathcal{E}_k$0. At inference, the reported implementation can optionally sharpen $\mathcal{E}_k$1 by selecting the top-$\mathcal{E}_k$2 expert for each sample and setting its weight to $\mathcal{E}_k$3 and all others to $\mathcal{E}_k$4, so that only one expert path is evaluated. In operational terms, this is the mechanism that couples input adaptivity to sparse inference.

4. Integration into the BEV multi-sensor fusion pipeline

The routed image feature $\mathcal{E}_k$5 is inserted into a standard camera-to-BEV projection module, described as being used “as in BEVFusion,” producing a camera BEV feature $\mathcal{E}_k$6. In parallel, the LiDAR branch processes raw point clouds through a voxel or pillar backbone to obtain $\mathcal{E}_k$7, and the radar branch produces $\mathcal{E}_k$8 (Xiang et al., 11 Aug 2025).

These modality-specific BEV features are aligned to the same grid size and fused either by concatenation followed by convolution,

$\mathcal{E}_k$9

or by element-wise summation when the dimensions match. The resulting BEV tensor is then passed to the detection head for classification and box regression.

The placement of the MoE only in the image-backbone stage is significant. It preserves the downstream fusion and prediction interfaces while allowing the image branch to adapt computationally and representationally to the input. The paper explicitly notes that sparse expert activation reduces the image-backbone cost from $$F^{(k)} = \mathcal{E}_k(F_{\mathrm{img}}) \in \mathbb{R}^{B \times C \times H' \times W'}.$$0 passes per sample to only one at inference, making the overall system nearly as efficient as the single-backbone counterpart. A plausible implication is that the framework seeks a compromise between modular specialization and deployment practicality, rather than maximizing expertization across the entire perception stack.

5. Training objective and load-balancing regularization

The optimization target combines a standard 3D detection loss with a load-balancing regularizer intended to prevent expert collapse. Let $$F^{(k)} = \mathcal{E}_k(F_{\mathrm{img}}) \in \mathbb{R}^{B \times C \times H' \times W'}.$$1 denote the detection loss, exemplified as focal loss for classification together with $$F^{(k)} = \mathcal{E}_k(F_{\mathrm{img}}) \in \mathbb{R}^{B \times C \times H' \times W'}.$$2 or smooth-$$F^{(k)} = \mathcal{E}_k(F_{\mathrm{img}}) \in \mathbb{R}^{B \times C \times H' \times W'}.$$3 losses for bounding box and orientation. If $$F^{(k)} = \mathcal{E}_k(F_{\mathrm{img}}) \in \mathbb{R}^{B \times C \times H' \times W'}.$$4 is the batch routing matrix, the paper defines

$$F^{(k)} = \mathcal{E}_k(F_{\mathrm{img}}) \in \mathbb{R}^{B \times C \times H' \times W'}.$$5

The balance loss is

$$F^{(k)} = \mathcal{E}_k(F_{\mathrm{img}}) \in \mathbb{R}^{B \times C \times H' \times W'}.$$6

which is described as being minimized when experts are used equally. The total loss is

$$F^{(k)} = \mathcal{E}_k(F_{\mathrm{img}}) \in \mathbb{R}^{B \times C \times H' \times W'}.$$7

with $$F^{(k)} = \mathcal{E}_k(F_{\mathrm{img}}) \in \mathbb{R}^{B \times C \times H' \times W'}.$$8 (Xiang et al., 11 Aug 2025).

The reported ablation directly ties this regularizer to system behavior. Without load balancing, CBDES MoE attains lower validation performance than the version with load balancing, which the paper interprets as evidence that balanced expert usage is important for stable training and effective specialization. In this formulation, the regularizer does not replace the task objective; it modulates expert utilization so that routing remains distributed rather than collapsing onto a narrow subset of experts.

6. Empirical results, interpretation, and stated boundaries

All reported experiments use the official nuScenes split of $$F^{(k)} = \mathcal{E}_k(F_{\mathrm{img}}) \in \mathbb{R}^{B \times C \times H' \times W'}.$$9 scenes and the standard metrics mean Average Precision (mAP) and nuScenes Detection Score (NDS). Under top-$1 \times 1$0 expert inference, the validation-set results are as follows (Xiang et al., 11 Aug 2025).

Model	mAP (%)	NDS (%)
BEVFusion–Swin Transformer	64.0	65.6
BEVFusion–ResNet	63.3	65.2
BEVFusion–ConvNeXt	61.6	65.2
BEVFusion–PVT	62.4	65.7
CBDES MoE (w/ LB)	65.6	69.8

The load-balancing ablation is reported separately:

Model	mAP (%)	NDS (%)
CBDES MoE (w/o LB)	63.4	65.8
CBDES MoE (w/ LB)	65.6	69.8

Relative to the strongest single-expert baselines, CBDES MoE yields a $1 \times 1$1-point increase in mAP and a $1 \times 1$2-point improvement in NDS. The paper attributes these gains to architectural heterogeneity, input-aware routing, and load balancing. This suggests that the reported performance improvement is not framed as the effect of any single backbone replacement alone, but as the result of the interaction among heterogeneous experts, SAR-based selection, and balanced expert usage.

The framework’s limits are also explicit. Although hierarchical decoupling is defined at the level of perception, fusion, and prediction modules, the current scope decouples only the perception module. Future extensions are described as potentially introducing MoE in fusion, for example through modality-specific experts, and in prediction, for example through task-specific experts. Consequently, CBDES MoE should be understood as a module-level MoE prototype centered on the perception branch of a BEV-based 3D perception system rather than as a complete expert-routed redesign of the full autonomous-driving stack.