BevAD: BEV Representation for Autonomous Driving

• BevAD is a paradigm that consolidates heterogeneous camera views into a unified BEV, reducing cross-view inconsistencies through explicit geometric transformation.
• It employs advanced modules like ROA-BEV and OA-BEV to enhance object feature learning before BEV projection, yielding measurable gains in mAP and NDS on nuScenes.
• Integrated with planning and control systems, BevAD demonstrates robust end-to-end performance while addressing challenges in domain adaptation, efficiency, and sensor fusion.

BevAD, shorthand for Bird’s-Eye-View for Autonomous Driving, denotes the use of a unified top-down representation in which multi-camera or multimodal observations are transformed into an ego-aligned spatial substrate for perception, prediction, planning, and control. In the literature considered here, the term refers both to a general paradigm and, in one 2026 work, to a specific lightweight end-to-end driving architecture named “BevAD”; in both usages, the central claim is that BEV reduces cross-view inconsistency, aligns inference with road geometry, and provides a common interface for downstream modules such as detection, tracking, motion forecasting, occupancy reasoning, and planning (Chen et al., 2024, Lu et al., 8 Aug 2025, Holtz et al., 16 Mar 2026).

1. Conceptual foundations and geometric formulation

BevAD systems are built around the observation that surround-view camera streams are naturally heterogeneous in perspective space but can be consolidated in a road-aligned coordinate frame. Representative formulations use camera intrinsics and extrinsics to lift image features into 3D and then pool them into a planar BEV grid. In Lift–Splat-style pipelines, a pixel $p=[u,v,1]^\top$ with depth $d$ is mapped by $X_{\mathrm{cam}}=dK^{-1}p$ and then transformed to ego or world coordinates before BEV discretization; BEVUDA states the same geometry as $X_{\mathrm{cam}}=dK^{-1}[u,v,1]^\top$, $$X_{\mathrm{ego}}=T_{\mathrm{cam}\rightarrow\mathrm{ego}}[X_{\mathrm{cam}};1]$$, followed by $$F_{\mathrm{BEV}}(x,y)=\sum_z F_{\mathrm{voxel}}(x,y,z)$$ (Liu et al., 2022). ME$^3$-BEV describes the resulting representation as translation equivariant, multi-view permutation invariant, and robust to calibration errors, while also emphasizing that it supplies a unified coordinate system for decision making (Lu et al., 8 Aug 2025).

The same literature also makes clear why this formulation is difficult. Camera-only BEV detectors inherit depth ambiguity, weather and illumination sensitivity, and view-transformation artifacts. ROA-BEV characterizes a recurring failure mode in which objects with strong similarity to the background in camera space are weakly encoded by the backbone and therefore degrade further after BEV projection; BEVUDA and BEVUDA++ generalize this point by arguing that distribution shift compounds across three spaces—2D image, 3D voxel, and BEV—rather than appearing only at the final detection head (Chen et al., 2024, Zhang et al., 17 Sep 2025). A plausible implication is that BevAD is best understood not as a single representation trick, but as a coupled geometric pipeline whose failure modes are distributed across sensing, lifting, fusion, and downstream reasoning.

2. BEV perception pipelines and object-centric feature learning

A canonical camera-only BevAD detector, as instantiated by ROA-BEV, contains an image backbone, an FPN, a depth estimation and view-transformation module in the BEVDepth family, a BEV encoder, and a BEV detection head for per-cell classification and 3D box regression. ROA-BEV augments this baseline with a 2D Region-Oriented Attention branch that consumes unfused multi-scale backbone features, processes each scale with an LKB stack comprising SE, two 7×7 basic blocks, ASPP, and a 7×7 DCN, fuses the resulting attention maps, and gates the FPN feature by element-wise multiplication before view transformation. The motivation is explicitly region-oriented supervision: the backbone is taught to emphasize regions where objects exist in the camera views, rather than distributing capacity across background regions (Chen et al., 2024).

On nuScenes val with CBGS, ROA-BEV improves BEVDepth (R50, 256×704) from mAP $0.351$ and NDS $0.475$ to mAP $0.361$ and NDS $d$0, with particularly notable gains in mAOE and mAVE while mATE and mASE remain nearly unchanged. Its ablations further show that multi-scale ROA is better than same-scale or FPN-only inputs, that overlap-based ROA supervision is stronger than binary labels, and that a 7×7 kernel is preferable to larger kernels, whose optimization degrades for 9×9 to 13×13 (Chen et al., 2024). These results suggest that objectness priors injected before lifting can be complementary to explicit depth supervision, rather than redundant with it.

OA-BEV addresses a related but distinct pathology: feature distortion and object blurring during image-to-BEV transformation. It introduces object-level depth supervision, a foreground selection stage using an FCOS detector, object-centric pseudo-3D features generated by projecting foreground pixels to pseudo-LiDAR, and deformable fusion into BEV through voxel deformable self-attention and depth deformable cross-attention. The method is intentionally plug-and-play and can be inserted into BEVDet and BEVFormer pipelines without altering their core view transformation or temporal aggregation (Chu et al., 2023).

Its quantitative evidence is consistent with that framing. On nuScenes val, BEVDet-Tiny improves from $d$1 mAP/NDS to $d$2 with OA-BEV, while BEVFormer-Base improves from $d$3 to $d$4; on the test set, BEVFormer* improves from $d$5 to $d$6. The ablations attribute gains separately to object-level depth supervision, pseudo-3D feature fusion, and depth feature fusion, with pseudo-3D features especially reducing translation error mATE by $d$7 (Chu et al., 2023). Together, OA-BEV and ROA-BEV define a recurrent BevAD design principle: strengthening BEV quality often requires modifying the image-domain feature learning stage before, or during, geometric transformation rather than relying only on a stronger BEV decoder.

3. From BEV perception to end-to-end planning and control

A major branch of BevAD research uses BEV not merely for detection but as the latent interface between perception and control. ME$d$8-BEV is a camera-only end-to-end DRL system that constructs a 200×200 BEV grid from six RGB views by a Lift–Splat encoder termed the Spatial-Semantic Aggregator, encodes navigation, road, vehicle, and BEV modalities, and then applies a Mamba-like temporal-aware fusion module with input-dependent state-space parameters before feeding the fused embedding into actor–critic PPO (Lu et al., 8 Aug 2025). Its temporal model is written as

$d$9

with $X_{\mathrm{cam}}=dK^{-1}p$0 and input-conditioned $X_{\mathrm{cam}}=dK^{-1}p$1. The objective operates in a POMDP with continuous steering and longitudinal control in $X_{\mathrm{cam}}=dK^{-1}p$2, and its reward couples collision penalties, speed tracking, and heading-aligned waypoint progress (Lu et al., 8 Aug 2025).

On CARLA, ME$X_{\mathrm{cam}}=dK^{-1}p$3-BEV reports strong reductions in collision rate relative to e2e-CLA. In low-density evaluation, collision rate falls from $X_{\mathrm{cam}}=dK^{-1}p$4 to $X_{\mathrm{cam}}=dK^{-1}p$5, timesteps increase from $X_{\mathrm{cam}}=dK^{-1}p$6 to $X_{\mathrm{cam}}=dK^{-1}p$7, and Driving Score increases from $X_{\mathrm{cam}}=dK^{-1}p$8 to $X_{\mathrm{cam}}=dK^{-1}p$9; in high density, collision rate falls from $X_{\mathrm{cam}}=dK^{-1}[u,v,1]^\top$0 to $X_{\mathrm{cam}}=dK^{-1}[u,v,1]^\top$1 and Driving Score rises from $X_{\mathrm{cam}}=dK^{-1}[u,v,1]^\top$2 to $X_{\mathrm{cam}}=dK^{-1}[u,v,1]^\top$3. Its ablations indicate that removing the BEV construction module markedly worsens safety, while removing temporal-aware fusion severely harms waypoint adherence, including a Town07 low-density Waypoint Distance increase to $X_{\mathrm{cam}}=dK^{-1}[u,v,1]^\top$4 from $X_{\mathrm{cam}}=dK^{-1}[u,v,1]^\top$5 (Lu et al., 8 Aug 2025). The paper’s interpretability decoder, which reconstructs semantic BEV maps from latent features, reinforces the view that BEV can remain structurally legible even inside an end-to-end control policy.

The 2026 architecture explicitly named BevAD re-examines a different question: which perception-to-planning interfaces actually scale in closed loop. It reports that open-loop architectural preferences do not necessarily transfer to Bench2Drive and identifies three interacting patterns: high-resolution perceptual representations, disentangled path-plus-speed planning tokens, and diffusion-based generative planning. Its own system retains a 100×100 BEV encoder based on BEVFormer and RADIO v2.5 with LoRA, but inserts a scene tokenizer that masks 20% of left and right BEV cells and patchifies the BEV grid before global cross-attention into the planning decoder (Holtz et al., 16 Mar 2026). The paper argues that this spatial bottleneck reduces causal confusion: without it, high-resolution global attention encourages planners to attend to distant or irrelevant tokens and overfit to spurious landmarks.

Closed-loop results on Bench2Drive are correspondingly strong. BevAD-S achieves Driving Score $X_{\mathrm{cam}}=dK^{-1}[u,v,1]^\top$6 and Success Rate $X_{\mathrm{cam}}=dK^{-1}[u,v,1]^\top$7, while BevAD-M reaches Driving Score $X_{\mathrm{cam}}=dK^{-1}[u,v,1]^\top$8 and Success Rate $X_{\mathrm{cam}}=dK^{-1}[u,v,1]^\top$9; selected baselines include UniAD at $$X_{\mathrm{ego}}=T_{\mathrm{cam}\rightarrow\mathrm{ego}}[X_{\mathrm{cam}};1]$$0 and SimLingo at $$X_{\mathrm{ego}}=T_{\mathrm{cam}\rightarrow\mathrm{ego}}[X_{\mathrm{cam}};1]$$1 for DS/SR. The same study also shows that compressing a high-resolution BEV is materially better than directly learning a low-resolution BEV with the same token count, and that diffusion plus a disentangled path-and-speed representation yields the highest success rates and the strongest data-scaling behavior (Holtz et al., 16 Mar 2026). This directly contradicts the common assumption that higher-resolution BEV features are uniformly beneficial for planning; within this literature, resolution is helpful for perception, but unbottlenecked access to it can degrade closed-loop robustness.

4. Domain adaptation, uncertainty, and robustness under corruption

Because BevAD pipelines traverse multiple geometric spaces, domain adaptation has been treated as a geometry-aware problem rather than a purely visual one. BEVUDA formulates unsupervised domain adaptation for multi-view camera-based 3D detection as a teacher–student problem over 2D image, 3D voxel, and BEV spaces. Its Depth-Aware Teacher uses target-domain LiDAR during training to replace or supplement predicted depth according to pixel-wise uncertainty from MC Dropout, while its Geometric-space Aligned Student projects 2D, voxel, and BEV features into a common embedding and aligns category-wise prototypes adversarially (Liu et al., 2022). The overall objective is

$$X_{\mathrm{ego}}=T_{\mathrm{cam}\rightarrow\mathrm{ego}}[X_{\mathrm{cam}};1]$$2

with $$X_{\mathrm{ego}}=T_{\mathrm{cam}\rightarrow\mathrm{ego}}[X_{\mathrm{cam}};1]$$3 and $$X_{\mathrm{ego}}=T_{\mathrm{cam}\rightarrow\mathrm{ego}}[X_{\mathrm{cam}};1]$$4.

Its reported gains are largest under severe shift. In Day→Night adaptation on nuScenes, BEVDepth (R101) achieves NDS $$X_{\mathrm{ego}}=T_{\mathrm{cam}\rightarrow\mathrm{ego}}[X_{\mathrm{cam}};1]$$5 and mAP $$X_{\mathrm{ego}}=T_{\mathrm{cam}\rightarrow\mathrm{ego}}[X_{\mathrm{cam}};1]$$6, whereas BEVUDA reaches NDS $$X_{\mathrm{ego}}=T_{\mathrm{cam}\rightarrow\mathrm{ego}}[X_{\mathrm{cam}};1]$$7 and mAP $$X_{\mathrm{ego}}=T_{\mathrm{cam}\rightarrow\mathrm{ego}}[X_{\mathrm{cam}};1]$$8; on Clear→Rainy, BEVUDA (R50) reaches NDS $$X_{\mathrm{ego}}=T_{\mathrm{cam}\rightarrow\mathrm{ego}}[X_{\mathrm{cam}};1]$$9 and mAP $$F_{\mathrm{BEV}}(x,y)=\sum_z F_{\mathrm{voxel}}(x,y,z)$$0 from a BEVDepth baseline of $$F_{\mathrm{BEV}}(x,y)=\sum_z F_{\mathrm{voxel}}(x,y,z)$$1. Ablations show that uncertainty-guided depth fusion is stronger than LiDAR-only replacement or confidence-guided fusion, and that distillation across more spaces yields monotonic gains (Liu et al., 2022). The method is deployment-oriented in a specific sense: LiDAR is used only during adaptation, while inference remains camera-only.

BEVUDA++ extends this line by replacing the earlier teacher with a Reliable Depth Teacher, introducing a Geometric Consistent Student, and adding Uncertainty-guided Exponential Moving Average:

$$F_{\mathrm{BEV}}(x,y)=\sum_z F_{\mathrm{voxel}}(x,y,z)$$2

with $$F_{\mathrm{BEV}}(x,y)=\sum_z F_{\mathrm{voxel}}(x,y,z)$$3 and $$F_{\mathrm{BEV}}(x,y)=\sum_z F_{\mathrm{voxel}}(x,y,z)$$4 (Zhang et al., 17 Sep 2025). The intent is to slow teacher updates when uncertainty is high and accelerate them when the student is reliable. On Day→Night, BEVUDA++ improves BEVUDA slightly, reaching NDS $$F_{\mathrm{BEV}}(x,y)=\sum_z F_{\mathrm{voxel}}(x,y,z)$$5 and mAP $$F_{\mathrm{BEV}}(x,y)=\sum_z F_{\mathrm{voxel}}(x,y,z)$$6 for R101; on Sunny→Foggy-3 it reaches $$F_{\mathrm{BEV}}(x,y)=\sum_z F_{\mathrm{voxel}}(x,y,z)$$7 with R50 from a BEVDepth baseline of $$F_{\mathrm{BEV}}(x,y)=\sum_z F_{\mathrm{voxel}}(x,y,z)$$8; and in continual fog it preserves markedly higher NDS than BEVDepth as density increases (Zhang et al., 17 Sep 2025). A plausible interpretation is that BevAD adaptation becomes increasingly effective when uncertainty is treated as a control signal for both geometric fusion and teacher dynamics, not merely as a confidence score.

Robustness to non-domain-shift disturbances has been addressed differently by RESBev, which reframes BEV robustness as latent semantic prediction rather than alignment. RESBev operates at the semantic BEV “Splat” stage of Lift–Splat–Shoot pipelines, predicts a clean BEV prior from the previous reconstructed BEV feature and ego-motion through a latent dynamic world model, and then reconstructs the current robust BEV by cross-attending from that prior to the current corrupted observation and prior history. It reports that the Splat stage is materially better than image-space Lift or task-space Shoot for recovery, and demonstrates large IoU gains across both natural corruptions and digital attacks (Zhuo et al., 10 Mar 2026).

The reported numbers are substantial. For LSS under seen corruptions, overall average IoU increases from $$F_{\mathrm{BEV}}(x,y)=\sum_z F_{\mathrm{voxel}}(x,y,z)$$9 to $^3$0, including PGD from $^3$1 to $^3$2 and Fog from $^3$3 to $^3$4; under unseen corruptions, the same baseline improves from $^3$5 to $^3$6 overall. Similar gains are reported for SimpleBEV, GaussianLSS, and FIERY, and the method often exceeds GraphBEV under corruption while leaving the underlying backbone unchanged (Zhuo et al., 10 Mar 2026). Within BevAD, this places robustness not only at the sensor or detector level, but at the level of BEV state evolution over time.

5. Efficiency, distillation, and multimodal expansion

BevAD has also expanded from heavy unified systems toward compact camera-only students. TinyBEV is a unified, camera-only BEV framework that distills a planning-oriented UniAD teacher into a student with a lightweight ResNet-18 per-view encoder, Lift–Splat–Shoot projection, a shared convolutional BEV backbone, and task heads for 3D detection, HD-map segmentation, motion forecasting, occupancy prediction, and goal-directed planning. Its distillation strategy combines feature-level BEV alignment, output-level imitation across all heads, and adaptive region-aware supervision that emphasizes dynamic agents, lanes, drivable boundaries, and intersections (Khan et al., 22 Sep 2025).

The method is explicitly positioned as an efficiency-preserving full-stack BevAD system. It reports a 28M-parameter backbone, a 78% parameter reduction relative to UniAD, and roughly 11 FPS, while preserving broad autonomy-stack coverage. On nuScenes, TinyBEV achieves 39.0 mAP for detection, 1.08 minADE for motion forecasting, and a 0.32 collision rate, while the full feature-plus-output distillation configuration improves markedly over no distillation: mAP rises from $^3$7 to $^3$8, minADE improves from $^3$9 to $0.351$0, L2@3s improves from $0.351$1 to $0.351$2, and collision drops from $0.351$3 to $0.351$4 (Khan et al., 22 Sep 2025). This suggests that a large fraction of “full-stack” BevAD capability can be retained in a camera-only deployment model when the supervision signal comes from a richer multimodal teacher.

A different extension appears in RESAR-BEV, which addresses BEV semantic segmentation with camera–radar fusion rather than camera-only perception. The framework uses six cameras and radar accumulated over six frames, a radar voxel encoder with max and attention pooling, learnable ground-proximity height offsets, and a progressive residual autoregressive refinement cascade comprising a Driver-Transformer and multiple Modifier-Transformers. Supervision is also progressive: the ground-truth BEV mask is decomposed offline into multi-scale token maps, and online training aligns each stage with its corresponding residual target before a final weighted Dice loss on the accumulated segmentation (Zeng et al., 10 May 2025).

On nuScenes, RESAR-BEV reports 54.0% mIoU across seven driving-scene categories at 14.6 FPS with 31.9M parameters. It also reports 40.8% vehicle mIoU at long range from 35–50 m, outperforming the compared baselines in that regime, and shows that removing radar, attention pooling, gating, or learnable height drift all degrades accuracy materially (Zeng et al., 10 May 2025). Although this work is not an end-to-end planner, it broadens BevAD’s scope by showing that BEV can remain the central representation even when the sensor stack becomes explicitly multimodal and the task becomes dense segmentation rather than detection or control.

6. Recurring limitations and open research directions

Across these works, several limitations recur. First, depth remains the principal bottleneck for camera-centric BevAD. ROA-BEV attributes missed detections to background-similar appearance before projection, OA-BEV to view-transformation blur and missing 3D structure, and BEVUDA-style methods to depth brittleness under domain shift, especially at night (Chen et al., 2024, Chu et al., 2023, Liu et al., 2022). Even when depth is improved through region attention, object-level supervision, LiDAR-assisted adaptation, or temporal prediction, the underlying geometric uncertainty is reduced rather than eliminated.

Second, stronger perception interfaces do not uniformly improve downstream behavior. The 2026 BevAD planner shows that high-resolution BEV grids with unrestricted global cross-attention can harm closed-loop performance through causal confusion, even when they remain useful for local perception; BEV-based augmentation improves closed-loop Driving Score from $0.351$5 to $0.351$6 when removed, while open-loop $0.351$7 remains roughly unchanged, illustrating a direct open-loop/closed-loop mismatch (Holtz et al., 16 Mar 2026). A common misconception is therefore that perception metrics or BEV fidelity transfer monotonically into driving robustness. The cited literature rejects that assumption.

Third, many of the strongest improvements introduce costs or dependencies that complicate deployment. ROA-BEV notes that large kernels and DCN increase memory and compute and require careful tuning, while not reporting FLOPs or latency (Chen et al., 2024). BEVUDA and BEVUDA++ require target LiDAR during adaptation, even though inference is camera-only (Liu et al., 2022, Zhang et al., 17 Sep 2025). RESBev assumes sequential history and reliable ego-motion, and RESAR-BEV’s radar fusion presupposes a multimodal platform rather than a minimal camera stack (Zhuo et al., 10 Mar 2026, Zeng et al., 10 May 2025). TinyBEV and the named BevAD architecture respond to this pressure from different directions—knowledge distillation in one case, architectural simplification and tokenizer bottlenecks in the other—but neither removes the broader trade-off between representational richness and operational efficiency (Khan et al., 22 Sep 2025, Holtz et al., 16 Mar 2026).

The future directions identified in the literature are correspondingly targeted. ROA-BEV points to kernel decomposition, sparse or dilated alternatives, temporal ROA, and softer attention labels such as instance-mask-derived priors (Chen et al., 2024). The BevAD planning paper calls for adaptive BEV masking, stronger command conditioning, and possibly reinforcement learning or VLM guidance to address residual causal confusion and red-light infractions (Holtz et al., 16 Mar 2026). ME$0.351$8-BEV emphasizes real-world deployment, sensor noise, and the simulation-to-reality gap (Lu et al., 8 Aug 2025). BEVUDA++ highlights robustness under more severe domain changes and better uncertainty calibration for teacher updates (Zhang et al., 17 Sep 2025). Taken together, these directions indicate that BevAD is moving from a representation-centric paradigm toward a broader systems question: how to construct BEV interfaces that are simultaneously geometrically faithful, temporally stable, robust under shift and corruption, and aligned with the decision structure of closed-loop driving.