面向安全端到端自动驾驶的风险感知占用

# Risk-Aware Occupancy for Safety-Oriented End-to-End Autonomous Driving
Source: [https://arxiv.org/html/2609.21470](https://arxiv.org/html/2609.21470)
Hengduo ZouYiren ZhaoBolin Gao††thanks:$†$ Corresponding author: Bolin Gao

###### Abstract

Sparse representation formulates the environment perception for the end\-to\-end driving system as a set of discrete elements like objects and lane lines\. This formulation meets safety risks in crowded, occluded scenes dealing with unstructured obstacles, uncertain regions, and intricate interactions\. In this paper, we propose a dense representation, risk\-aware occupancy, to characterize planning\-relevant risks in an explicit and uniform manner\. It jointly encodes global scene occupancy, map\-derived traffic constraints, and future dynamic agent occupancy into a unified BEV map\. The unified BEV map captures the risk evidence for trajectory planning in both spatial and temporal dimensions\. We design an E2E network, ROIDrive, to realize risk\-aware occupancy\. It predicts risk\-aware occupancy with an independent branch and injects it into planning queries for safety\-oriented trajectory generation\. In addition, to quantify the safety problem, we introduce RiskOcc4D\-nuScenes built upon nuscenes and occ3d\-nuscenes\. Our risk\-aware occupancy yields relative open\-loop collision reductions of52\.9%under the UniAD metric and35\.0%under the ST\-P3 metric on nuScenes\.

1School of Vehicle and Mobility, Tsinghua University

2The Hong Kong University of Science and Technology \(Guangzhou\)

## Introduction

Safety is one of the core problems blocking the wide application of both L2 and L4 autonomous driving\. Current end\-to\-end methods have made great progress on the safety problem, but the sparse representation makes it hard to completely and explicitly tackle the planning\-relevant risks\. When the environment is encoded into discrete agents and maps elements, a natural safety risk arises in dealing with unstructured elements and occluded regions\. In contrast, a dense representation could describe the scene as a continuous spatial field\. Then the planner can explicitly check whether candidate trajectories are feasible, and measure the conflict risk around them\. In this paper, we propose risk\-aware occupancy as the dense representation and make it flexible to help existing e2e frameworks become safety\-oriented\.

![Refer to caption](https://arxiv.org/html/2609.21470v1/nips_teaser.png)Figure 1:Motivation and design of risk\-aware occupancy\. Left: Limitations of sparse agent/map detection and naive 3D occupancy, and human\-inspired requirements for dense representations covering geometry, drivability, traffic rules and vulnerable road users\. Middle: Comparison between conventional separated E2E pipelines and our unified prediction\-planning framework, which constructs global dynamic risk\-aware occupancy by encoding collision and traffic regulation risks\. Right: Collision\-rate benchmarks on UniAD and ST\-P3 metrics, verifying the safety improvement of the proposed risk\-aware occupancy representation\.By diving into the detailed scenes of safe driving, we summarize two key attributes for the dense representation\. As illustrated in Figure[1](https://arxiv.org/html/2609.21470#Sx1.F1), first, it should continuously distinguish drivable from non\-drivable space, which has an arbitrary shape, in the time dimension\. Then, it should evaluate spatial conflicts by predicting how these conflicts evolve over time\. These insights follow the intuitive human driving principles, where a driver will determine whether the surround enviroment could let the vehicle safely pass before considering the fine\-grained categories\.\([Mecheri et al\. 2023](https://arxiv.org/html/2609.21470#bib.bib33);[Papakostopoulos et al\. 2017](https://arxiv.org/html/2609.21470#bib.bib34)\)\. For example, a relatively occupied height indicates whether a region is physically drivable for safety\. A map structure could define the admissible motion space in law and avoid dangerous driving\. Furthermore, the future motion could reveal emerging conflicts to prevent possible collisions\. After that, fine\-grained semantic information could help deal with traffic rules and hazard severity\([Tavakoli et al\. 2023](https://arxiv.org/html/2609.21470#bib.bib36);[Tavakoli and Heydarian 2022](https://arxiv.org/html/2609.21470#bib.bib35)\)\. Therefore, we propose the risk\-aware occupancy map as the dense perception representation, which could explicitly model 3D geometry and spatial interaction and be interpretable for safety inspection\. It encodes the drivable area, spatial conflicts, and future motion into one unified 3d space\. For safety\-oriented driving, it also directly computes the planning\-relevant risk beyond the geometric and semantic information\.

Inspired by\([Chen et al\. 2024](https://arxiv.org/html/2609.21470#bib.bib19)\)\. The risk\-aware occupancy is built on a unified bev risk field, where each cell could model a comparable risk score\. The risk score is computed from object geometry, safety\-relevant semantic priors, map constraints, and future motion cues\. With the unified field, the ego vehicle can encounter risk directly queryable by the planner\. Furthermore, projecting planning\-relevant 3D structure into BEV reduces redundant vertical representation and matches the spatial feature layout used by modern E2E networks\. Therefore, the risk\-aware occupancy can serve both as a supervisory target for perception and as an intermediate feature for planning\. It aligns the perception objective with downstream decisions and provides a direct interface for safety inspection, failure localization, and visual debugging\.

We design a risk\-aware occupancy enhanced instance\-centric E2E \(ROIDrive\) to prove the priority of risk\-aware occupancy for an end\-to\-end framework\. The risk\-aware occupancy is predicted by an independent branch and fused into planning\. The integration is seamless for any kind of end\-to\-end network design by replacing the sparse representation with our risk\-aware occupancy map\. Besides, to ensure efficient performance, we also tokenize the dense risk field by sliding\-window sampling and inject it into the agent queries via cross\-attention, staying orthogonal to the sparse tokens\.

To fully study the safety problem, we also present a new datasetRiskOcc4D\-nuScenesfrom Occ3D\-nuScenes\. The dataset is built through four automated annotation pipelines providing a reliable risk coverage under occlusion and distant point\-cloud sparsity, which incorporates road\-related and temporal cues\. We conduct ablation on multiple 3D occupancy architectures to predict risk\-aware occupancy, establishing a benchmark for representation learnability and model compatibility\. By only replacing the oracle risk maps with our predicted risk\-aware occupancy, we can separate the intrinsic value of the representation from prediction error and validate its practical utility for downstream planning\.

The main contributions of this work are summarized as follows:

1. 1\.Risk\-aware occupancy representation\.We introduce a safety\-oriented dense BEV representation that unifies geometry, safety\-relevant semantics, map constraints, and future motion into an explicit planning\-relevant risk field\.
2. 2\.ROIDrive E2E framework\.We design ROIDrive, an E2E driving framework with an independent risk\-aware occupancy branch seamlessly integrated at the planning stage, and validate it throughout the complete perception\-to\-planning pipeline\.
3. 3\.RiskOcc4D\-nuScenes dataset\.We construct the RiskOcc4D\-nuScenes dataset using four automated annotation pipelines and instantiate risk\-aware occupancy in the end\-to\-end framework to verify its effectiveness and quantify its performance improvements\.

## Related Work

### Occupancy Prediction Methods

Recently, occupancy prediction has advanced rapidly along three axes\.Lightweight architectures: FlashOcc\([Yu et al\. 2023](https://arxiv.org/html/2609.21470#bib.bib8)\), TPVFormer\([Huang et al\. 2023](https://arxiv.org/html/2609.21470#bib.bib9)\), SparseOcc\([Tang et al\. 2024](https://arxiv.org/html/2609.21470#bib.bib11)\), FastOcc\([Hou et al\. 2024](https://arxiv.org/html/2609.21470#bib.bib18)\), and Octree\-Occ\([Lu et al\. 2023](https://arxiv.org/html/2609.21470#bib.bib22)\)reduce cost via BEV extraction, tri\-plane decomposition, sparse encoding, and octree reconstruction, respectively\.Performance optimization: FB\-Occ\([Li et al\. 2023b](https://arxiv.org/html/2609.21470#bib.bib4)\)fuses forward\-backward projection\([Huang et al\. 2021](https://arxiv.org/html/2609.21470#bib.bib1);[Li et al\. 2022](https://arxiv.org/html/2609.21470#bib.bib13)\)with multi\-task learning; F\-OCC\([Zhao et al\. 2024](https://arxiv.org/html/2609.21470#bib.bib25)\)extends DCNv4\([Xiong et al\. 2024](https://arxiv.org/html/2609.21470#bib.bib27)\)to 3D; OccTransformer\([Liu et al\. 2024](https://arxiv.org/html/2609.21470#bib.bib20)\), OccFormer\([Zhang et al\. 2023b](https://arxiv.org/html/2609.21470#bib.bib3)\), VoxelFormer\([Li et al\. 2023a](https://arxiv.org/html/2609.21470#bib.bib5)\), and MambaOcc\([Tian et al\. 2024](https://arxiv.org/html/2609.21470#bib.bib23)\)leverage advanced attention or state\-space architectures\.Cross\-domain integration: NeRF\-based methods\([Pan et al\. 2023](https://arxiv.org/html/2609.21470#bib.bib21);[Zhang et al\. 2023a](https://arxiv.org/html/2609.21470#bib.bib28);[Huang et al\. 2024b](https://arxiv.org/html/2609.21470#bib.bib10);[Mildenhall et al\. 2021](https://arxiv.org/html/2609.21470#bib.bib24)\), Gaussian splatting variants\([Huang et al\. 2024c](https://arxiv.org/html/2609.21470#bib.bib12);[Huang et al\. 2024a](https://arxiv.org/html/2609.21470#bib.bib31);[Gan et al\. 2024](https://arxiv.org/html/2609.21470#bib.bib17);[Kerbl et al\. 2023](https://arxiv.org/html/2609.21470#bib.bib26)\), world models\([Zheng et al\. 2025](https://arxiv.org/html/2609.21470#bib.bib6)\), and LLM\-driven approaches\([Xu et al\. 2025](https://arxiv.org/html/2609.21470#bib.bib15);[Wei et al\. 2024](https://arxiv.org/html/2609.21470#bib.bib16)\)further enrich occupancy representations from complementary perspectives\.

![Refer to caption](https://arxiv.org/html/2609.21470v1/roi_architecture.png)Figure 2:ROIDrive architecture\. A BEVFormer backbone yields the BEV featureBB; the risk\-aware occupancy \(RiskOcc\) branch regresses a normalized risk field via a Bi\-Mamba2 block and a class\-to\-risk lookup table \(optionally fused with an SDMap\+ prior\), tokenizes it by sliding\-window sampling, and injects it into the queries via cross\-attention\. Temporal Query Consistency \(TQC\) then gates the queries by cross\-frame reliability before a VAE generative planner decodes multimodal trajectories\.
### Auxiliary Perception Supervision in End\-to\-End Driving Models

Recent end\-to\-end driving models widely adopt multi\-task perception auxiliary losses to regularize shared BEV features\. Typical branches include BEV 2D/3D detection, semantic segmentation, vectorized lane prediction, and dense 3D/4D occupancy forecasting\. For instance, VADv2\([Jiang et al\. 2026](https://arxiv.org/html/2609.21470#bib.bib47)\)and DiffusionDriveV2\([Zou et al\. 2025](https://arxiv.org/html/2609.21470#bib.bib48)\)use BEV detection to constrain agent localization; HENet\+\+\([Xia et al\. 2025](https://arxiv.org/html/2609.21470#bib.bib42)\)adds map segmentation for lane\-compliant trajectories; and OccVLA\([Liu et al\. 2025](https://arxiv.org/html/2609.21470#bib.bib49)\)and IR\-WM\([Mei et al\. 2025](https://arxiv.org/html/2609.21470#bib.bib50)\)exploit volumetric occupancy for vertical obstacle geometry, whereas the lightweight DrivoR\([Kirby et al\. 2026](https://arxiv.org/html/2609.21470#bib.bib51)\)drops dense occupancy to cut cost at the expense of spatial risk cues\. Yet all such supervision optimizes geometric or semantic objectives decoupled from trajectory ranking, and full 4D occupancy remains costly\. We instead propose risk\-aware occupancy, a compressed BEV risk signal that embeds trajectory\-consistent spatiotemporal risk at low computational cost\.

![Refer to caption](https://arxiv.org/html/2609.21470v1/pipelines.png)Figure 3:The four automated annotation pipelines building RiskOcc4D\-nuScenes, fusing static, regulatory, drivable\-area, and dynamic\-future cues into the global and dynamic risk\-aware occupancy maps\.

## Risk\-Aware Occupancy Perception

Conventional 3D semantic occupancy builds high\-entropy voxel representationsH⁡\(𝒱\)H\(\\mathcal\{V\}\)with fine\-grained categories; though complete, they carry substantial task\-irrelevant redundancyHred​\(𝒱\)H\_\{\\mathrm\{red\}\}\(\\mathcal\{V\}\)that degrades risk\-information transfer for E2E driving\.

Risk\-aware occupancy prediction is formulated as planning\-oriented minimal sufficient statistic extraction:

R∗=arg⁡maxR⁡I⁡\(R,Tplan\),s\.t\.​dim\(R\)≪dim\(𝒱\),R^\{\*\}=\\arg\\max\_\{R\}\\;I\(R;\\,T\_\{\\mathrm\{plan\}\}\),\\quad\\text\{s\.t\.\}\\;\\;\\dim\(R\)\\ll\\dim\(\\mathcal\{V\}\),\(1\)whereI⁡\(R,Tplan\)I\(R;\\,T\_\{\\mathrm\{plan\}\}\)is the mutual information between risk representationRRand the optimal trajectoryTplanT\_\{\\mathrm\{plan\}\}\.

Moreover, conventional 3D occupancy is over\-complete: its semantic paradigm cannot uniformly quantify collision threats and regulatory constraints \(*e\.g\.*, lane markings, stop lines\)\. Risk\-aware occupancy encodes both risk types and resolves conflicts via maximum aggregation over collision and regulatory terms\. To realize this formulation, dual\-axis reduction is performed:

- •Spatial:Only the critical height intervalHcrit∈\[0​m,2\.2​m\]H\_\{\\mathrm\{crit\}\}\\in\[0\\;\\mathrm\{m\},\\,2\.2\\;\\mathrm\{m\}\]\(the vehicle\-relevant collision zone\) is retained, discarding non\-threatening airspace\.
- •Semantic:Fine\-grained labels are replaced by object height as the primary risk quantifier; vulnerable road users \(VRUs\) retain semantic priors with elevated risk weights\.

A hierarchical risk quantification framework is constructed over the Occ3D\-nuScenes perception domainΩ=\{\(x,y\)∣x,y∈\[−40m,40m\]\}\\Omega=\\\{\(x,y\)\\mid x,y\\in\[\-40\\;\\mathrm\{m\},\\,40\\;\\mathrm\{m\}\]\\\}\. Critical heights are nonlinearly mapped to a normalized risk scoreZnorm∈\[0,1\]Z\_\{\\mathrm\{norm\}\}\\in\[0,1\]\. To separate risk modes, static global obstacle risk is bounded toVriskglobal∈\[0\.2,0\.7\]V\_\{\\mathrm\{risk\}\}^\{\\mathrm\{global\}\}\\in\[0\.2,\\,0\.7\]and moving objects to a dedicated branch withVriskdyn∈\[0\.5,1\.0\]V\_\{\\mathrm\{risk\}\}^\{\\mathrm\{dyn\}\}\\in\[0\.5,\\,1\.0\]\. Their deliberate overlap enforces a strict priority hierarchy—dynamic collision\>\>static obstacle\>\>regulatory risk—so dynamic threats receive planning priority while gradient conflict in multi\-task optimization is mitigated\.

### ROIDrive Framework

Design rationale\.As illustrated in Figure[2](https://arxiv.org/html/2609.21470#Sx2.F2), ROIDrive builds on the query\-based generative E2E paradigm\([Zheng et al\. 2024](https://arxiv.org/html/2609.21470#bib.bib32);[Jiang et al\. 2023](https://arxiv.org/html/2609.21470#bib.bib39)\), where a BEVFormer backbone yields a BEV featureB∈ℝC×H×WB\\in\\mathbb\{R\}^\{C\\times H\\times W\}, a small fixed set of agent and map queries abstracts the scene, and a conditional VAE planner decodes multimodal trajectories\. This compression causes two structural weaknesses:*representation sparsity*, as the dense, temporally evolving risk structure of the BEV plane is left implicit, and*token sparsity and unreliability*, as missed detections erase risk while false\-positive and flickering queries inject spurious motion intent\. ROIDrive thus pursues two goals: \(1\) to introduce a perception modality that is independently predicted and then fused into the E2E pipeline, and \(2\) to densify the representation and calibrate token reliability\. Both are met by two modules at the motion\-prediction stage: a risk\-aware occupancy \(RiskOcc\) branch and a Temporal Query Consistency \(TQC\) module\.

Risk\-aware occupancy branch\.The BEV feature is channel\-reduced, modeled by a Bi\-Mamba2 state\-space block that captures long\-range dependencies at𝒪⁡\(N\)\\mathcal\{O\}\(N\)cost, and upsampled\. A lookup table converts the stable class predictions to a normalized risk fieldR∈\[0,0\.7\]R\\in\[0,0\.7\]; after optional fusion with an SDMap\+ prior,RRis tokenized by sliding\-window sampling and injected into the queries via cross\-attention\. The field is dense and orthogonal to the queries: the queries encode object identity and location, while the field encodes the current and near\-future risk at each location\. It also provides a detection\-free fallback for missed hazards, sinceRRis regressed directly fromBB, and anticipates risk through long\-range scanning and future supervision\.

Temporal Query Consistency\.TQC calibrates query reliability before motion self\-attention without learnable parameters\. It caches previous\-frame agent states, aligns them to the current ego frame via CAN\-bus ego\-motion, and matches each query to its nearest predecessor\. A consistency weight fuses feature cosine similarity with a distance\-decayed spatial score, with a fallback for newly appearing objects and a lower bound against hard deletion\. The weights form a pairwise geometric\-mean gate mapped to a log\-odds biasβ\\betathat modulates attention through a stop\-gradientsg⁡\(⋅\)\\mathrm\{sg\}\(\\cdot\):Attn⁡\(Q,K,V\)=softmax⁡\(Q​K⊤/d\+sg⁡\(β\)\)​V\\mathrm\{Attn\}\(Q,K,V\)=\\mathrm\{softmax\}\(QK^\{\\top\}/\\sqrt\{d\}\+\\mathrm\{sg\}\(\\beta\)\)V\. Consistent tokens are amplified and spurious or flickering ones driven toward zero attention; being inference\-only and gradient\-free, TQC is lightweight and falls back to the base model on the first frame\.

Generative planning\.The instance tokens, densified by RiskOcc, reweighted by TQC, and initialized at the ego slot from historical ego states\([Li et al\. 2023b](https://arxiv.org/html/2609.21470#bib.bib4)\), are encoded by a VAE into a latent Gaussianp⁡\(z∣T\)∼𝒩⁡\(μi,σi\)p\(z\\mid T\)\\sim\\mathcal\{N\}\(\\mu\_\{i\},\\sigma\_\{i\}\)\([Zheng et al\. 2024](https://arxiv.org/html/2609.21470#bib.bib32)\); a sampledzzis decoded by a GRU and an MLP into multimodal ego and agent trajectories\.

## RiskOcc4D\-nuScenes Dataset

Conventional 3D occupancy datasets suffer from spatial redundancy, semantics misaligned with E2E planning, the absence of a unified collision\-regulatory risk encoding, and missing temporal risk annotations\. We address these limitations withRiskOcc4D\-nuScenes, built upon nuScenes and Occ3D\-nuScenes—the first 4D risk\-aware occupancy benchmark for E2E autonomous driving\. Four automated pipelines unify static global risk, traffic regulatory risk, and dynamic future risk into a single framework while remaining fully compatible with existing 3D occupancy algorithms\.

### Task Definition

Combining the avoidance\-priority principle with Bayesian risk decision theory, we instantiate the abstract optimal representationR∗R^\{\*\}in Eq\. \([1](https://arxiv.org/html/2609.21470#Sx3.E1)\) by minimizing the bottleneck risk that aligns perception with planning:

ℒ⁡\(Tplan∣Z\)=𝔼⁡\[max⁡\(ℓcol​\(Tplan,O\),ℓreg​\(Tplan,M\)\)∣Z\],\\mathcal\{L\}\(T\_\{\\mathrm\{plan\}\}\\mid Z\)=\\mathbb\{E\}\\\!\\left\[\\max\\\!\\left\(\\ell\_\{\\mathrm\{col\}\}\(T\_\{\\mathrm\{plan\}\},O\),\\;\\ell\_\{\\mathrm\{reg\}\}\(T\_\{\\mathrm\{plan\}\},M\)\\right\)\\mid Z\\right\],\(2\)whereZZis the multi\-view input,\(ℓcol,O\)\(\\ell\_\{\\mathrm\{col\}\},O\)the collision risk and obstacle set, and\(ℓreg,M\)\(\\ell\_\{\\mathrm\{reg\}\},M\)the regulatory risk and traffic\-rule constraints\. Themax\\maxoperator enforces hard\-constraint alignment, avoiding accumulation overflow and inter\-term gradient conflict, thereby providing the supervisory signal for the risk\-aware occupancy maps in Section[Dataset Construction Pipelines](https://arxiv.org/html/2609.21470#Sx4.SSx2)\.

Input\.Six surround\-view camera sequences under nuScenes calibration: intrinsics\{Ki\}\\\{K\_\{i\}\\\}, extrinsics\{\[Ri∣ti\]\}\\\{\[R\_\{i\}\\mid t\_\{i\}\]\\\}, andNNconsecutive frames\{Ii,t∈ℝHi×Wi×3\}\\\{I\_\{i,t\}\\in\\mathbb\{R\}^\{H\_\{i\}\\times W\_\{i\}\\times 3\}\\\}\(i=1,…,6i\{=\}1,\\dots,6;t=1,…,Nt\{=\}1,\\dots,N\)\.

Table 1:Feasibility benchmark for predicting risk\-aware occupancy with representative 3D occupancy backbones\.BaselinesReleaseMetricsGlobalDynamicInfer\. SpeedmIoU↑\\uparrowMAE↓\\downarrowMAE@15↓\\downarrowMAE@30↓\\downarrowSSIM↑\\uparrowMask\-MAE↓\\downarrowFPS↑\\uparrowPanoOcc\-S\([Wang et al\. 2024](https://arxiv.org/html/2609.21470#bib.bib14)\)CVPR 2414\.540\.1210\.0220\.0790\.1310\.7326\.0PanoOcc\([Wang et al\. 2024](https://arxiv.org/html/2609.21470#bib.bib14)\)CVPR 2414\.710\.1210\.0220\.0790\.1310\.7311\.9OccProphet\([Chen et al\. 2025](https://arxiv.org/html/2609.21470#bib.bib46)\)ICLR 2517\.390\.0990\.0170\.0670\.3550\.5240\.7BEVDet4DOcc\([Huang and Huang 2022](https://arxiv.org/html/2609.21470#bib.bib2)\)\-19\.710\.0930\.0150\.0620\.3731\.8569\.5ViewFormerOcc\([Li et al\. 2024](https://arxiv.org/html/2609.21470#bib.bib44)\)ECCV 2421\.380\.0930\.0150\.0600\.3790\.73419\.0FO\(BEVDET\)\([Yu et al\. 2023](https://arxiv.org/html/2609.21470#bib.bib8)\)ICCV 2322\.030\.0880\.0140\.0580\.4240\.68266\.2BEVFormer\-Occ\([Li et al\. 2022](https://arxiv.org/html/2609.21470#bib.bib13)\)ECCV 2223\.760\.0860\.0140\.0570\.4370\.56511\.1CVTOcc\([Ye et al\. 2024](https://arxiv.org/html/2609.21470#bib.bib41)\)ECCV 2424\.430\.0940\.0160\.0630\.3850\.6232\.1FO\(Stereo4D\)\([Yu et al\. 2023](https://arxiv.org/html/2609.21470#bib.bib8)\)ICCV 2325\.140\.0820\.0110\.0520\.4430\.66410\.4DHDOcc\-S\([Wu et al\. 2025](https://arxiv.org/html/2609.21470#bib.bib37)\)ICRA 2525\.450\.0920\.0150\.0610\.3990\.61018\.7ProtoOcc\([Kim et al\. 2025](https://arxiv.org/html/2609.21470#bib.bib45)\)AAAI 2528\.580\.0820\.0120\.0530\.4220\.5897\.1FB\-Occ\([Li et al\. 2023b](https://arxiv.org/html/2609.21470#bib.bib4)\)ICCV 2332\.960\.0750\.0100\.0480\.4800\.67534\.1DHDOcc\-M\([Wu et al\. 2025](https://arxiv.org/html/2609.21470#bib.bib37)\)ICRA 2534\.150\.0790\.0110\.0500\.4550\.5593\.5Output\.Ego\-centric BEV risk\-aware occupancy maps: a global mapR​Mt∈ℝH×WRM\_\{t\}\\in\\mathbb\{R\}^\{H\\times W\}, where occupancy encodes spatial location and risk values quantify hazard attention carrying collision and regulatory information, and a dynamic mapR​MdRM\_\{d\}that fuses future stepsΔ​t\\Delta tto encode motion\-prediction uncertainty\. They provide three targets: \(1\) discrete global risk classification, \(2\) continuous global risk regression, and \(3\) dynamic temporal risk prediction\.

### Dataset Construction Pipelines

Our four\-stage automated framework generates the global and dynamic risk\-aware occupancy maps \(Figure[3](https://arxiv.org/html/2609.21470#Sx2.F3)\), compressing redundant 3D representations while enriching label completeness\.

#### Static Annotation Pipeline \(Pipeline\-1\)

This pipeline extracts static risk priors from Occ3D\-nuScenes via 3D\-to\-BEV reduction\. For eachxx\-yycell, vertical marginalization within the critical height rangeHcrit∈\[0,2\.2\]​mH\_\{\\mathrm\{crit\}\}\\in\[0,\\,2\.2\]\\,\\mathrm\{m\}projects semantic and geometric cues onto the BEV plane, where the driving risk attentionVriskV\_\{\\mathrm\{risk\}\}is a joint function of local height distribution and semantic category priors\. To suppress quantization artifacts and smooth jagged boundaries, we refine the initial risk field𝐑init\\mathbf\{R\}\_\{\\mathrm\{init\}\}via morphological dilation\-erosion and a custom convolution\-kernel infilling:𝐑^0=ℱ⁡\(\(𝐑init⊕𝐊d\)⊖𝐊e\),\\hat\{\\mathbf\{R\}\}\_\{0\}=\\mathcal\{F\}\\\!\\left\(\(\\mathbf\{R\}\_\{\\mathrm\{init\}\}\\oplus\\mathbf\{K\}\_\{d\}\)\\ominus\\mathbf\{K\}\_\{e\}\\right\),where⊕\\oplusand⊖\\ominusdenote dilation and erosion with structuring elements𝐊d\\mathbf\{K\}\_\{d\}and𝐊e\\mathbf\{K\}\_\{e\}, andℱ⁡\(⋅\)\\mathcal\{F\}\(\\cdot\)propagates neighboring context to repair sparse regions\.

#### Lane Line Annotation Pipeline \(Pipeline\-2\)

This pipeline extracts an ego\-centric80​m×80​m80\\,\\mathrm\{m\}\{\\times\}80\\,\\mathrm\{m\}semantic map from nuScenes and injects regulatory priors by assigningVrisk=0\.1V\_\{\\mathrm\{risk\}\}=0\.1to road\-marking cells \(crosswalks, stop lines, lane dividers\)\. This explicit rule encoding prunes the trajectory hypothesis space, reduces decision entropy, and sets a low\-risk regulatory baseline in the hierarchical risk framework\.

#### Drivable Road Annotation Pipeline \(Pipeline\-3\)

To counter missing drivable\-area labels and LiDAR\-sparsity noise in Occ3D\-nuScenes, high\-fidelity drivable masks from nuScenes HD maps are fused with the Pipeline\-1/2 outputs through deterministic spatial fusion, completing undefined areas and removing outliers \(Figure[3](https://arxiv.org/html/2609.21470#Sx2.F3)\)\. For occluded blind\-spot cells with unobservable occupancy, a non\-informative prior of0\.050\.05—below regulatory and collision risk yet strictly positive—encodes epistemic uncertainty, guiding cautious attention toward unobserved regions until later viewpoints resolve a definite ground truth\. Fusing Pipelines 1–3 yields the global risk\-aware occupancy map, stored as both classification and regression labels\.

#### Dynamic Annotation Pipeline \(Pipeline\-4\)

This pipeline builds the dynamic risk potential field from observable temporal occupancy rather than unreliable kinematic fitting\. Moving objects \(v\>0\.5​m/sv\>0\.5\\,\\mathrm\{m/s\}\) are tracked across frames to record actual occupied positions over the next2\.5​s2\.5\\,\\mathrm\{s\}at0\.5​s0\.5\\,\\mathrm\{s\}intervals\. Risk attention is inversely proportional to the time step—near\-future moments carry higher attention, following the “urgent\-near, relaxed\-far” defensive\-driving logic—yielding an initial field𝐑^0dyn\\hat\{\\mathbf\{R\}\}\_\{0\}^\{\\mathrm\{dyn\}\}\. A spatial Gaussian diffusion𝐑dyn=𝒢σ∗𝐑^0dyn\\mathbf\{R\}^\{\\mathrm\{dyn\}\}=\\mathcal\{G\}\_\{\\sigma\}\*\\hat\{\\mathbf\{R\}\}\_\{0\}^\{\\mathrm\{dyn\}\}\(kernel𝒢σ\\mathcal\{G\}\_\{\\sigma\},∗\*spatial convolution\) produces a continuous field with Lipschitz\-continuous variation, modeling risk attenuation toward free space\. Per Section[Risk\-Aware Occupancy Perception](https://arxiv.org/html/2609.21470#Sx3), dynamic objects takeVriskdyn∈\[0\.5,1\.0\]V\_\{\\mathrm\{risk\}\}^\{\\mathrm\{dyn\}\}\\in\[0\.5,\\,1\.0\]and are stored separately from static labels to avoid multi\-task interference\.

## Experiments

We evaluate multiple 3D occupancy backbones on RiskOcc4D\-nuScenes and its downstream utility for E2E driving\.

### Experimental Setup

#### Implementation details

All occupancy baselines use their default configurations with only the output heads replaced: the classification head retains each method’s original supervision, while both regression heads use the Huber loss\([Huber 1964](https://arxiv.org/html/2609.21470#bib.bib43)\)\. Training runs on multiple NVIDIA A100 GPUs; inference and benchmarking use an Intel i9\-13900K CPU with an NVIDIA RTX 4090 GPU\.

#### Evaluation Metrics

We adopt a four\-fold protocol\.\(1\) Classificationuses mIoU \(mean Jaccard index\)\([Everingham et al\. 2010](https://arxiv.org/html/2609.21470#bib.bib29)\)for risk\-category alignment;\(2\) Regressionuses MAE \(L1L\_\{1\}\), with MAE@15/MAE@30 restricted to ego\-centric15×1515\{\\times\}15and30×30​m230\{\\times\}30\\,\\text\{m\}^\{2\}regions to emphasize the safety\-critical near field;\(3\) Structural fidelityuses SSIM\([Wang et al\. 2004](https://arxiv.org/html/2609.21470#bib.bib30)\)for luminance, contrast, and structural consistency beyond pointwise accuracy\.\(4\) Dynamic predictionuses Mask\-MAE, which counters the extreme imbalance from dominant background zeros via a binary maskMi=𝟏\[Pi\>τorGi\>τ\]M\_\{i\}=\\mathbf\{1\}\[P\_\{i\}\>\\tau\\;\\text\{or\}\\;G\_\{i\}\>\\tau\]\(τ=0\.1\\tau=0\.1\):

Mask\-MAE=1∑iMi​∑i=1NMi​\|Pi−Gi\|,\\text\{Mask\-MAE\}=\\tfrac\{1\}\{\\sum\_\{i\}M\_\{i\}\}\\sum\_\{i=1\}^\{N\}M\_\{i\}\|P\_\{i\}\-G\_\{i\}\|,\(3\)restricting evaluation to the supportmax⁡\(Pi,Gi\)\>τ\\max\(P\_\{i\},G\_\{i\}\)\>\\tauand improving discriminability\.

![Refer to caption](https://arxiv.org/html/2609.21470v1/gain-exp.png)Figure 4:Architecture of the RiskOcc\-GenAD fusion experiment: the predicted risk\-aware occupancy map is sampled by a sliding window and fused with GenAD’s instance tokens via cross\-attention, entering as an additional input while GenAD’s protocols stay unchanged\.

### Benchmarking on RiskOcc4D\-nuScenes

We migrate a broad set of 3D occupancy methods to RiskOcc4D\-nuScenes—PanoOcc\([Wang et al\. 2024](https://arxiv.org/html/2609.21470#bib.bib14)\), OccProphet\([Chen et al\. 2025](https://arxiv.org/html/2609.21470#bib.bib46)\), BEVDet4DOcc\([Huang and Huang 2022](https://arxiv.org/html/2609.21470#bib.bib2)\), ViewFormerOcc\([Li et al\. 2024](https://arxiv.org/html/2609.21470#bib.bib44)\), FlashOcc\(FO\)\([Yu et al\. 2023](https://arxiv.org/html/2609.21470#bib.bib8)\), BEVFormer\-Occ\([Li et al\. 2022](https://arxiv.org/html/2609.21470#bib.bib13)\), CVTOcc\*\([Ye et al\. 2024](https://arxiv.org/html/2609.21470#bib.bib41)\), ProtoOcc\([Kim et al\. 2025](https://arxiv.org/html/2609.21470#bib.bib45)\), FB\-Occ\([Li et al\. 2023b](https://arxiv.org/html/2609.21470#bib.bib4)\), and DHDOcc\([Wu et al\. 2025](https://arxiv.org/html/2609.21470#bib.bib37)\)—with results in Table[1](https://arxiv.org/html/2609.21470#Sx4.T1)\.

Table 2:Collision rate \(%\) comparison among E2E methods\. Ours = GenAD\([Zheng et al\. 2024](https://arxiv.org/html/2609.21470#bib.bib32)\)\+\+RiskOcc Map\. P\. = progressive training\.MethodsST\-P3 Metrics↓\\downarrowUniAD Metrics↓\\downarrow1 s2 s3 sAvg\.1 s2 s3 sAvg\.ST\-P30\.230\.621\.270\.71\-\-\-\-OccNet0\.210\.591\.370\.72\-\-\-\-VAD\-Tiny0\.210\.350\.580\.38\-\-\-\-VAD\-Base0\.070\.170\.410\.220\.010\.300\.950\.42FusionAD0\.250\.130\.250\.21\-\-\-\-UniAD\-\-\-\-0\.050\.170\.710\.31PPAD \(P\.\)0\.080\.120\.380\.190\.020\.200\.930\.38PPAD0\.070\.150\.360\.190\.030\.220\.730\.33PARADrive\-\-\-\-0\.070\.250\.600\.30AD\-MLP0\.170\.180\.240\.20\-\-\-\-GenAD0\.050\.290\.870\.400\.060\.150\.320\.17Ours0\.000\.140\.640\.260\.020\.040\.190\.08Global risk\-aware occupancy\.DHDOcc\-M\([Wu et al\. 2025](https://arxiv.org/html/2609.21470#bib.bib37)\)leads in mIoU \(34\.15\), indicating strong risk\-category discrimination, while FB\-Occ\([Li et al\. 2023b](https://arxiv.org/html/2609.21470#bib.bib4)\)attains the best regression at all scales \(MAE 0\.075, MAE@15 0\.010, MAE@30 0\.048\) and the highest SSIM \(0\.480\), best preserving the ground\-truth spatial gradients\.

Dynamic risk\-aware occupancy\.DHDOcc\-M achieves the best Mask\-MAE \(0\.559\) via its decoupled temporal architecture, with BEVFormer\-Occ close behind \(0\.565\), underscoring the value of temporal attention\. BEVDet4DOcc is an outlier \(1\.856\), likely from distribution shift in its dynamic head; all others fall within 0\.559–0\.734\. The generally limited performance highlights the challenge of multi\-step risk propagation\.

Inference efficiency\.FO\(BEVDET\)\([Yu et al\. 2023](https://arxiv.org/html/2609.21470#bib.bib8)\)reaches 66\.2 FPS without temporal history but with limited accuracy, whereas FB\-Occ\([Li et al\. 2023b](https://arxiv.org/html/2609.21470#bib.bib4)\)offers the best speed\-accuracy trade\-off \(34\.1 FPS\) alongside the strongest global metrics\.

Table 3:Ablation of intermediate representations on collision rate \(%\)\. Best inbold\.InputsST\-P3 Metrics↓\\downarrowUniAD Metrics↓\\downarrow1 s2 s3 sAvg\.1 s2 s3 sAvg\.RiskOcc Map0\.000\.140\.640\.260\.020\.040\.190\.083D\-Occ0\.080\.290\.740\.370\.090\.180\.330\.20Binary Map0\.140\.390\.900\.480\.110\.210\.390\.24Gaussian Map0\.270\.510\.860\.550\.210\.320\.370\.33
### End\-to\-End Autonomous Driving Integration

#### Experimental Design

We integrate the risk\-aware occupancy map into GenAD\([Zheng et al\. 2024](https://arxiv.org/html/2609.21470#bib.bib32)\), an E2E planner encoding dynamic agents and road topology as instance and map tokens\. Such discrete object\-level representations are spatially incomplete, capturing only detected targets while lacking explicit encoding of dense spatial risk, long\-tail obstacles, and ambiguous occupied regions—gaps that our dense, continuous risk distribution complements\. Let𝒵inst\+map\\mathcal\{Z\}\_\{\\text\{inst\+map\}\}and𝒵risk\\mathcal\{Z\}\_\{\\text\{risk\}\}denote the joint instance\-and\-map and the risk representations\. Our central hypothesis is a positive conditional mutual information gain:

Δ​I=I⁡\(𝒜safe;𝒵risk∣𝒵inst\+map\)\>0,\\Delta I=I\\\!\\left\(\\mathcal\{A\}\_\{\\text\{safe\}\};\\,\\mathcal\{Z\}\_\{\\text\{risk\}\}\\mid\\mathcal\{Z\}\_\{\\text\{inst\+map\}\}\\right\)\>0,\(4\)i\.e\., the risk representation carries safety\-relevant information not captured by instance and map tokens\. The risk map is sampled via a sliding window and fused with GenAD’s instance tokens through cross\-attention \(Figure[4](https://arxiv.org/html/2609.21470#Sx5.F4)\), allowing each instance token to attend to surrounding grid\-level risk\. GenAD’s original training and evaluation protocols remain unchanged and the risk map enters solely as an additional input, ensuring all gains are attributable to the risk representation\.

Table 4:ROIDrive ablation 1\. M1: Risk\-Aware Occupancy Perception; M2: ego\-token prior initialization; M3: TQC \(Temporal Query Consistency\)\.MethodsUniAD MetricsST\-P3 MetricsM1M2M3L2 \(m\)↓\\downarrowCollision Rate \(%\)↓\\downarrowL2 \(m\)↓\\downarrowCollision Rate \(%\)↓\\downarrow1 s2 s3 sAvg\.1 s2 s3 sAvg\.1 s2 s3 sAvg\.1 s2 s3 sAvg\.✓0\.290\.510\.780\.520\.080\.070\.210\.120\.380\.851\.520\.920\.070\.070\.640\.26✓✓0\.160\.310\.560\.340\.080\.100\.270\.150\.210\.581\.220\.670\.080\.160\.840\.36✓✓✓0\.160\.320\.570\.350\.000\.040\.210\.080\.210\.591\.260\.690\.000\.140\.860\.33
#### Comparative Results

Table[2](https://arxiv.org/html/2609.21470#Sx5.T2)reports collision rates on the nuScenes validation set\. Ours substantially reduces collisions over GenAD under both metrics:↓\\downarrow52\.9% on UniAD\([Hu et al\. 2023](https://arxiv.org/html/2609.21470#bib.bib40)\)\(0\.17%→\\to0\.08% Avg\.\) and↓\\downarrow35\.0% on ST\-P3\([Hu et al\. 2022](https://arxiv.org/html/2609.21470#bib.bib38)\)\(0\.40%→\\to0\.26% Avg\.\), consistently across all horizons and largest at the 2 s horizon \(UniAD↓\\downarrow73\.3%; ST\-P3↓\\downarrow51\.7%\), where accumulated trajectory uncertainty is high and dense spatial risk priors are most valuable\. Ours also attains zero collisions at 1 s under ST\-P3 and the second\-lowest under UniAD \(0\.02%\), achieving the lowest average collision rate under the UniAD metric among all compared methods\.

#### Ablation Study

Each ablation replaces the RiskOcc Map with an alternative spatial representation while keeping the fusion module and all other settings fixed \(Table[3](https://arxiv.org/html/2609.21470#Sx5.T3)\)\.

3D\-Occ\.Conventional 3D semantic voxels are adaptively pooled to the RiskOcc Map size and fused identically\. Gains are mixed—slightly better than baseline under ST\-P3 \(0\.37 vs\. 0\.40\) but worse under UniAD \(0\.20 vs\. 0\.17\): although raw voxels carry high mutual information, adaptive pooling lowers the effective information rate and planning\-irrelevant semantic redundancy adds noise\. By distilling safety\-critical information at construction time, RiskOcc attains higher task\-relevant density in a lower\-dimensional BEV, with consistent gains under both protocols \(0\.26 / 0\.08\)\.

Binary Map\.Collapsing continuous risk to a single bit \(risk = 1, no risk = 0\) erases ordinal relationships among threat levels\. The resulting rates \(ST\-P3 Avg\. 0\.48; UniAD Avg\. 0\.24\) are far worse than RiskOcc \(0\.26 / 0\.08\) and even degrade over the no\-map baseline \(0\.40 / 0\.17\), showing that a gradient\-free binary signal misleads the planner\. Hence risk stratification, rather than mere spatial occupancy, constitutes the core mechanism behind the safety gains\.

Gaussian Map\.As a control, a random Gaussian map of identical size but carrying no risk semantics yields the worst rates across all metrics \(ST\-P3 Avg\. 0\.55; UniAD Avg\. 0\.33\), below the no\-map baseline, ruling out gains from the added cross\-attention parameters and confirming that RiskOcc’s benefit stems from semantically meaningful risk encoding\.

Together, these ablations rule out network capacity, high\-dimensional semantics, and binary occupancy as alternative explanations, providing evidence consistent with—though not a direct measurement of—the conditional mutual information gainΔ​I\>0\\Delta I\>0in Eq\. \([4](https://arxiv.org/html/2609.21470#Sx5.E4)\)\.

### Planning Performance of ROIDrive

Overall performance\.On the nuScenes\([Caesar et al\. 2020](https://arxiv.org/html/2609.21470#bib.bib7)\)validation set, ROIDrive attains an average L2 error and collision rate of0\.350\.35m /0\.08%0\.08\\%under the UniAD\([Hu et al\. 2023](https://arxiv.org/html/2609.21470#bib.bib40)\)protocol and0\.690\.69m /0\.33%0\.33\\%under the ST\-P3\([Hu et al\. 2022](https://arxiv.org/html/2609.21470#bib.bib38)\)protocol, with the lowest collision rate under the UniAD protocol among competing methods\.

Table 5:ROIDrive ablation 2 on the RiskOcc\-only model \(without ego\-token initialization or TQC\)\. C1: Bi\-Mamba2; C2: SDMap\+ Fusion\.ComponentsUniAD Collision Rate↓\\downarrowST\-P3 Collision Rate↓\\downarrowC1C21 s2 s3 sAvg\.1 s2 s3 sAvg\.✓0\.080\.110\.300\.160\.080\.180\.840\.37✓0\.050\.120\.220\.130\.060\.270\.540\.29✓✓0\.080\.070\.210\.120\.070\.070\.640\.26Ablation: framework modules\.Table[4](https://arxiv.org/html/2609.21470#Sx5.T4)incrementally adds the modules\. RiskOcc alone \(M1\) sets a dense risk prior and a safe planning baseline; ego\-token initialization \(M2\) adds the ego motion prior and sharply lowers average L2 \(UniAD→0\.340\.52\\\!\\to\\\!0\.34, ST\-P3→0\.670\.92\\\!\\to\\\!0\.67\) at marginal collision cost; TQC \(M3\) then reduces collisions relative to the ego\-initialized variant \(UniAD0\.15%→0\.08%0\.15\\%\\\!\\to\\\!0\.08\\%, ST\-P30\.36%→0\.33%0\.36\\%\\\!\\to\\\!0\.33\\%\) by down\-weighting unreliable queries while keeping L2 nearly unchanged, achieving the intended safety\-accuracy balance\.

Ablation: occupancy components\.Built on the RiskOcc\-only model without ego\-token initialization or TQC, Table[5](https://arxiv.org/html/2609.21470#Sx5.T5)isolates the two components of the branch\. The Bi\-Mamba2 backbone \(C1\) provides dynamic avoidance from onboard vision, whereas the SDMap\+ prior \(C2\) fills static blind spots and most reduces the33s collision rate; enabling both yields the lowest average collision rate \(UniAD0\.12%0\.12\\%, ST\-P30\.26%0\.26\\%\), confirming that dynamic perception and static priors are complementary\.

## Conclusion

This paper introduced risk\-aware occupancy as a new perception representation for E2E autonomous driving\. We constructed RiskOcc4D\-nuScenes, fully compatible with Occ3D\-nuScenes, under a standardized multi\-level evaluation protocol\. We validated our 3D occupancy methods and demonstrated their compatibility and transferability\. By fusing the predicted risk\-aware occupancy map into GenAD\([Zheng et al\. 2024](https://arxiv.org/html/2609.21470#bib.bib32)\)via sliding\-window sampling, we yielded collision\-rate reductions of 52\.9% \(UniAD\) and 35\.0% \(ST\-P3\)\. We further designed the ROIDrive, an E2E framework coupling an independent risk\-aware occupancy branch with temporal query consistency for safety\-oriented planning\.

## References

- Caesaret al\.\(2020\)H\. Caesar, V\. Bankiti, A\. H\. Lang, S\. Vora, V\. E\. Liong, Q\. Xu, A\. Krishnan, Y\. Pan, G\. Baldan, and O\. BeijbomNuscenes: a multimodal dataset for autonomous driving\.InProceedings of the IEEE/CVF conference on computer vision and pattern recognition,pp\. 11621–11631\.Cited by:[Planning Performance of ROIDrive](https://arxiv.org/html/2609.21470#Sx5.SSx4.p1.1)\.
- Chenet al\.\(2024\)J\. Chen, W\. Zhong, B\. Gao, Y\. Liu, H\. Zou, J\. Liu, Y\. Lu, J\. Huang, and Z\. ZhongRisk occupancy: a new and efficient paradigm through vehicle\-road\-cloud collaboration\.arXiv preprint arXiv:2408\.07367\.Cited by:[Introduction](https://arxiv.org/html/2609.21470#Sx1.p3.1)\.
- Chenet al\.\(2025\)J\. Chen, H\. Xu, Y\. Wang, and L\. ChauOccprophet: pushing efficiency frontier of camera\-only 4d occupancy forecasting with observer\-forecaster\-refiner framework\.arXiv preprint arXiv:2502\.15180\.Cited by:[Table 1](https://arxiv.org/html/2609.21470#Sx4.T1.1.6.1),[Benchmarking on RiskOcc4D\-nuScenes](https://arxiv.org/html/2609.21470#Sx5.SSx2.p1.1)\.
- Everinghamet al\.\(2010\)M\. Everingham, L\. Van Gool, C\. K\. Williams, J\. Winn, and A\. ZissermanThe pascal visual object classes \(voc\) challenge\.International journal of computer vision88,pp\. 303–338\.Cited by:[Evaluation Metrics](https://arxiv.org/html/2609.21470#Sx5.SSx1.SSSx2.p1.1)\.
- Ganet al\.\(2024\)W\. Gan, F\. Liu, H\. Xu, N\. Mo, and N\. YokoyaGaussianOcc: fully self\-supervised and efficient 3d occupancy estimation with gaussian splatting\.arXiv preprint arXiv:2408\.11447\.Cited by:[Occupancy Prediction Methods](https://arxiv.org/html/2609.21470#Sx2.SSx1.p1.1)\.
- Houet al\.\(2024\)J\. Hou, X\. Li, W\. Guan, G\. Zhang, D\. Feng, Y\. Du, X\. Xue, and J\. PuFastOcc: accelerating 3d occupancy prediction by fusing the 2d bird’s\-eye view and perspective view\.arXiv preprint arXiv:2403\.02710\.Cited by:[Occupancy Prediction Methods](https://arxiv.org/html/2609.21470#Sx2.SSx1.p1.1)\.
- Huet al\.\(2022\)S\. Hu, L\. Chen, P\. Wu, H\. Li, J\. Yan, and D\. TaoSt\-p3: end\-to\-end vision\-based autonomous driving via spatial\-temporal feature learning\.InEuropean Conference on Computer Vision,pp\. 533–549\.Cited by:[Comparative Results](https://arxiv.org/html/2609.21470#Sx5.SSx3.SSSx2.p1.1),[Planning Performance of ROIDrive](https://arxiv.org/html/2609.21470#Sx5.SSx4.p1.1)\.
- Huet al\.\(2023\)Y\. Hu, J\. Yang, L\. Chen, K\. Li, C\. Sima, X\. Zhu, S\. Chai, S\. Du, T\. Lin, W\. Wang, L\. Lu, X\. Jia, Q\. Liu, J\. Dai, Y\. Qiao, and H\. LiPlanning\-oriented autonomous driving\.InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,Cited by:[Comparative Results](https://arxiv.org/html/2609.21470#Sx5.SSx3.SSSx2.p1.1),[Planning Performance of ROIDrive](https://arxiv.org/html/2609.21470#Sx5.SSx4.p1.1)\.
- Huanget al\.\(2021\)J\. Huang, G\. Huang, Z\. Zhu, Y\. Ye, and D\. DuBevdet: high\-performance multi\-camera 3d object detection in bird\-eye\-view\.arXiv preprint arXiv:2112\.11790\.Cited by:[Occupancy Prediction Methods](https://arxiv.org/html/2609.21470#Sx2.SSx1.p1.1)\.
- Huang and Huang \(2022\)J\. Huang and G\. HuangBevdet4d: exploit temporal cues in multi\-camera 3d object detection\.arXiv preprint arXiv:2203\.17054\.Cited by:[Table 1](https://arxiv.org/html/2609.21470#Sx4.T1.1.7.1.1),[Benchmarking on RiskOcc4D\-nuScenes](https://arxiv.org/html/2609.21470#Sx5.SSx2.p1.1)\.
- Huanget al\.\(2024a\)Y\. Huang, A\. Thammatadatrakoon, W\. Zheng, Y\. Zhang, D\. Du, and J\. LuProbabilistic gaussian superposition for efficient 3d occupancy prediction\.arXiv preprint arXiv:2412\.04384\.Cited by:[Occupancy Prediction Methods](https://arxiv.org/html/2609.21470#Sx2.SSx1.p1.1)\.
- Huanget al\.\(2024b\)Y\. Huang, W\. Zheng, B\. Zhang, J\. Zhou, and J\. LuSelfocc: self\-supervised vision\-based 3d occupancy prediction\.InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,pp\. 19946–19956\.Cited by:[Occupancy Prediction Methods](https://arxiv.org/html/2609.21470#Sx2.SSx1.p1.1)\.
- Huanget al\.\(2023\)Y\. Huang, W\. Zheng, Y\. Zhang, J\. Zhou, and J\. LuTri\-perspective view for vision\-based 3d semantic occupancy prediction\.InProceedings of the IEEE/CVF conference on computer vision and pattern recognition,pp\. 9223–9232\.Cited by:[Occupancy Prediction Methods](https://arxiv.org/html/2609.21470#Sx2.SSx1.p1.1)\.
- Huanget al\.\(2024c\)Y\. Huang, W\. Zheng, Y\. Zhang, J\. Zhou, and J\. LuGaussianFormer: scene as gaussians for vision\-based 3d semantic occupancy prediction\.arXiv preprint arXiv:2405\.17429\.Cited by:[Occupancy Prediction Methods](https://arxiv.org/html/2609.21470#Sx2.SSx1.p1.1)\.
- Huber \(1964\)P\. J\. HuberRobust estimation of a location parameter\.The Annals of Mathematical Statistics35\(1\),pp\. 73–101\.Cited by:[Implementation details](https://arxiv.org/html/2609.21470#Sx5.SSx1.SSSx1.p1.1)\.
- Jianget al\.\(2026\)B\. Jiang, S\. Chen, H\. Gao, B\. Liao, Q\. Zhang, W\. Liu, and X\. WangVADv2: end\-to\-end vectorized autonomous driving via probabilistic planning\.InInternational Conference on Learning Representations,External Links:[Link](https://openreview.net/forum?id=0a4dA6eUHN)Cited by:[Auxiliary Perception Supervision in End\-to\-End Driving Models](https://arxiv.org/html/2609.21470#Sx2.SSx2.p1.1)\.
- Jianget al\.\(2023\)B\. Jiang, S\. Chen, Q\. Xu, B\. Liao, J\. Chen, H\. Zhou, Q\. Zhang, W\. Liu, C\. Huang, and X\. WangVad: vectorized scene representation for efficient autonomous driving\.InProceedings of the IEEE/CVF International Conference on Computer Vision,pp\. 8340–8350\.Cited by:[ROIDrive Framework](https://arxiv.org/html/2609.21470#Sx3.SSx1.p1.1)\.
- Kerblet al\.\(2023\)B\. Kerbl, G\. Kopanas, T\. Leimkühler, and G\. Drettakis3d gaussian splatting for real\-time radiance field rendering\.\.ACM Trans\. Graph\.42\(4\),pp\. 139–1\.Cited by:[Occupancy Prediction Methods](https://arxiv.org/html/2609.21470#Sx2.SSx1.p1.1)\.
- Kimet al\.\(2025\)J\. Kim, C\. Kang, D\. Lee, S\. Choi, and J\. W\. ChoiProtoocc: accurate, efficient 3d occupancy prediction using dual branch encoder\-prototype query decoder\.InProceedings of the AAAI Conference on Artificial Intelligence,Vol\.39,pp\. 4284–4292\.Cited by:[Table 1](https://arxiv.org/html/2609.21470#Sx4.T1.1.14.1),[Benchmarking on RiskOcc4D\-nuScenes](https://arxiv.org/html/2609.21470#Sx5.SSx2.p1.1)\.
- Kirbyet al\.\(2026\)E\. Kirby, A\. Boulch, Y\. Xu, Y\. Yin, G\. Puy, É\. Zablocki, A\. Bursuc, S\. Gidaris, R\. Marlet, F\. Bartoccioni, A\. Cao, N\. Samet, T\. Vu, and M\. CordDriving on registers\.InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,External Links:[Link](https://arxiv.org/abs/2601.05083)Cited by:[Auxiliary Perception Supervision in End\-to\-End Driving Models](https://arxiv.org/html/2609.21470#Sx2.SSx2.p1.1)\.
- Liet al\.\(2024\)J\. Li, X\. He, C\. Zhou, X\. Cheng, Y\. Wen, and D\. ZhangViewFormer: exploring spatiotemporal modeling for multi\-view 3d occupancy perception via view\-guided transformers\.External Links:2405\.04299,[Link](https://arxiv.org/abs/2405.04299)Cited by:[Table 1](https://arxiv.org/html/2609.21470#Sx4.T1.1.8.1),[Benchmarking on RiskOcc4D\-nuScenes](https://arxiv.org/html/2609.21470#Sx5.SSx2.p1.1)\.
- Liet al\.\(2023a\)Y\. Li, Z\. Yu, C\. Choy, C\. Xiao, J\. M\. Alvarez, S\. Fidler, C\. Feng, and A\. AnandkumarVoxformer: sparse voxel transformer for camera\-based 3d semantic scene completion\.InProceedings of the IEEE/CVF conference on computer vision and pattern recognition,pp\. 9087–9098\.Cited by:[Occupancy Prediction Methods](https://arxiv.org/html/2609.21470#Sx2.SSx1.p1.1)\.
- Liet al\.\(2022\)Z\. Li, W\. Wang, H\. Li, E\. Xie, C\. Sima, T\. Lu, Y\. Qiao, and J\. DaiBevformer: learning bird’s\-eye\-view representation from multi\-camera images via spatiotemporal transformers\.InEuropean conference on computer vision,pp\. 1–18\.Cited by:[Occupancy Prediction Methods](https://arxiv.org/html/2609.21470#Sx2.SSx1.p1.1),[Table 1](https://arxiv.org/html/2609.21470#Sx4.T1.1.10.1),[Benchmarking on RiskOcc4D\-nuScenes](https://arxiv.org/html/2609.21470#Sx5.SSx2.p1.1)\.
- Liet al\.\(2023b\)Z\. Li, Z\. Yu, D\. Austin, M\. Fang, S\. Lan, J\. Kautz, and J\. M\. AlvarezFb\-occ: 3d occupancy prediction based on forward\-backward view transformation\.arXiv preprint arXiv:2307\.01492\.Cited by:[Occupancy Prediction Methods](https://arxiv.org/html/2609.21470#Sx2.SSx1.p1.1),[ROIDrive Framework](https://arxiv.org/html/2609.21470#Sx3.SSx1.p4.1),[Table 1](https://arxiv.org/html/2609.21470#Sx4.T1.1.15.1),[Benchmarking on RiskOcc4D\-nuScenes](https://arxiv.org/html/2609.21470#Sx5.SSx2.p1.1),[Benchmarking on RiskOcc4D\-nuScenes](https://arxiv.org/html/2609.21470#Sx5.SSx2.p2.1),[Benchmarking on RiskOcc4D\-nuScenes](https://arxiv.org/html/2609.21470#Sx5.SSx2.p4.1)\.
- Liuet al\.\(2024\)J\. Liu, S\. Zhang, C\. Kong, W\. Zhang, Y\. Wu, Y\. Ding, B\. Xu, R\. Ming, D\. Wei, and X\. LiuOccTransformer: improving bevformer for 3d camera\-only occupancy prediction\.arXiv preprint arXiv:2402\.18140\.Cited by:[Occupancy Prediction Methods](https://arxiv.org/html/2609.21470#Sx2.SSx1.p1.1)\.
- Liuet al\.\(2025\)R\. Liu, L\. Kong, D\. Li, and H\. ZhaoOccVLA: vision\-language\-action model with implicit 3D occupancy supervision\.arXiv preprint arXiv:2509\.05578\.External Links:[Document](https://dx.doi.org/10.48550/arXiv.2509.05578),[Link](https://arxiv.org/abs/2509.05578)Cited by:[Auxiliary Perception Supervision in End\-to\-End Driving Models](https://arxiv.org/html/2609.21470#Sx2.SSx2.p1.1)\.
- Luet al\.\(2023\)Y\. Lu, X\. Zhu, T\. Wang, and Y\. MaOctreeOcc: efficient and multi\-granularity occupancy prediction using octree queries\.arXiv preprint arXiv:2312\.03774\.Cited by:[Occupancy Prediction Methods](https://arxiv.org/html/2609.21470#Sx2.SSx1.p1.1)\.
- Mecheriet al\.\(2023\)S\. Mecheri, F\. Mars, and R\. LobjoisInfluence of continuous edge\-line delineation on drivers’ lateral positioning in curves: a gaze\-steering approach\.Ergonomics67,pp\. 1–24\.External Links:[Document](https://dx.doi.org/10.1080/00140139.2023.2226844)Cited by:[Introduction](https://arxiv.org/html/2609.21470#Sx1.p2.1)\.
- Meiet al\.\(2025\)J\. Mei, Y\. Yang, X\. Yang, L\. Wen, J\. Lv, B\. Shi, and Y\. LiuVision\-centric 4D occupancy forecasting and planning via implicit residual world models\.arXiv preprint arXiv:2510\.16729\.External Links:[Document](https://dx.doi.org/10.48550/arXiv.2510.16729),[Link](https://arxiv.org/abs/2510.16729)Cited by:[Auxiliary Perception Supervision in End\-to\-End Driving Models](https://arxiv.org/html/2609.21470#Sx2.SSx2.p1.1)\.
- Mildenhallet al\.\(2021\)B\. Mildenhall, P\. P\. Srinivasan, M\. Tancik, J\. T\. Barron, R\. Ramamoorthi, and R\. NgNerf: representing scenes as neural radiance fields for view synthesis\.Communications of the ACM65\(1\),pp\. 99–106\.Cited by:[Occupancy Prediction Methods](https://arxiv.org/html/2609.21470#Sx2.SSx1.p1.1)\.
- Panet al\.\(2023\)M\. Pan, L\. Liu, J\. Liu, P\. Huang, L\. Wang, S\. Zhang, S\. Xu, Z\. Lai, and K\. YangUniocc: unifying vision\-centric 3d occupancy prediction with geometric and semantic rendering\.arXiv preprint arXiv:2306\.09117\.Cited by:[Occupancy Prediction Methods](https://arxiv.org/html/2609.21470#Sx2.SSx1.p1.1)\.
- Papakostopouloset al\.\(2017\)V\. Papakostopoulos, N\. Marmaras, and D\. NathanaelThe “field of safe travel” revisited: interpreting driving behaviour performance through a holistic approach\.Transport Reviews37\(6\),pp\. 695–714\.External Links:ISSN 0144\-1647,[Document](https://dx.doi.org/https%3A//doi.org/10.1080/01441647.2017.1289992),[Link](https://www.sciencedirect.com/science/article/pii/S0144164722003221)Cited by:[Introduction](https://arxiv.org/html/2609.21470#Sx1.p2.1)\.
- Tanget al\.\(2024\)P\. Tang, Z\. Wang, G\. Wang, J\. Zheng, X\. Ren, B\. Feng, and C\. MaSparseocc: rethinking sparse latent representation for vision\-based semantic occupancy prediction\.InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,pp\. 15035–15044\.Cited by:[Occupancy Prediction Methods](https://arxiv.org/html/2609.21470#Sx2.SSx1.p1.1)\.
- Tavakoli and Heydarian \(2022\)A\. Tavakoli and A\. HeydarianThe impact of surrounding road objects and conditions on drivers abrupt heart rate changes\.External Links:2205\.11014,[Link](https://arxiv.org/abs/2205.11014)Cited by:[Introduction](https://arxiv.org/html/2609.21470#Sx1.p2.1)\.
- Tavakoliet al\.\(2023\)A\. Tavakoli, N\. Lai, V\. Balali, and A\. HeydarianHow are drivers stress levels and emotions associated with the driving context, a naturalistic study\.Journal of Transport & Health31,pp\. 101649\.External Links:ISSN 2214\-1405,[Document](https://dx.doi.org/https%3A//doi.org/10.1016/j.jth.2023.101649)Cited by:[Introduction](https://arxiv.org/html/2609.21470#Sx1.p2.1)\.
- Tianet al\.\(2024\)Y\. Tian, S\. Bai, Z\. Luo, Y\. Wang, Y\. Lv, and F\. WangMambaocc: visual state space model for bev\-based occupancy prediction with local adaptive reordering\.arXiv preprint arXiv:2408\.11464\.Cited by:[Occupancy Prediction Methods](https://arxiv.org/html/2609.21470#Sx2.SSx1.p1.1)\.
- Wanget al\.\(2024\)Y\. Wang, Y\. Chen, X\. Liao, L\. Fan, and Z\. ZhangPanoocc: unified occupancy representation for camera\-based 3d panoptic segmentation\.InProceedings of the IEEE/CVF conference on computer vision and pattern recognition,pp\. 17158–17168\.Cited by:[Table 1](https://arxiv.org/html/2609.21470#Sx4.T1.1.4.1.1),[Table 1](https://arxiv.org/html/2609.21470#Sx4.T1.1.5.1),[Benchmarking on RiskOcc4D\-nuScenes](https://arxiv.org/html/2609.21470#Sx5.SSx2.p1.1)\.
- Wanget al\.\(2004\)Z\. Wang, A\. C\. Bovik, H\. R\. Sheikh, and E\. P\. SimoncelliImage quality assessment: from error visibility to structural similarity\.IEEE transactions on image processing13\(4\),pp\. 600–612\.Cited by:[Evaluation Metrics](https://arxiv.org/html/2609.21470#Sx5.SSx1.SSSx2.p1.1)\.
- Weiet al\.\(2024\)J\. Wei, S\. Yuan, P\. Li, Q\. Hu, Z\. Gan, and W\. DingOccllama: an occupancy\-language\-action generative world model for autonomous driving\.arXiv preprint arXiv:2409\.03272\.Cited by:[Occupancy Prediction Methods](https://arxiv.org/html/2609.21470#Sx2.SSx1.p1.1)\.
- Wuet al\.\(2025\)Y\. Wu, Z\. Yan, Z\. Wang, X\. Li, L\. Hui, and J\. YangDeep height decoupling for precise vision\-based 3d occupancy prediction\.External Links:2409\.07972,[Link](https://arxiv.org/abs/2409.07972)Cited by:[Table 1](https://arxiv.org/html/2609.21470#Sx4.T1.1.13.1.1),[Table 1](https://arxiv.org/html/2609.21470#Sx4.T1.1.16.1.1),[Benchmarking on RiskOcc4D\-nuScenes](https://arxiv.org/html/2609.21470#Sx5.SSx2.p1.1),[Benchmarking on RiskOcc4D\-nuScenes](https://arxiv.org/html/2609.21470#Sx5.SSx2.p2.1)\.
- Xiaet al\.\(2025\)Z\. Xia, Z\. Lin, Y\. Wang, and M\. YangHENet\+\+: hybrid encoding and multi\-task learning for 3d perception and end\-to\-end autonomous driving\.External Links:2511\.07106,[Link](https://arxiv.org/abs/2511.07106)Cited by:[Auxiliary Perception Supervision in End\-to\-End Driving Models](https://arxiv.org/html/2609.21470#Sx2.SSx2.p1.1)\.
- Xionget al\.\(2024\)Y\. Xiong, Z\. Li, Y\. Chen, F\. Wang, X\. Zhu, J\. Luo, W\. Wang, T\. Lu, H\. Li, Y\. Qiao,et al\.Efficient deformable convnets: rethinking dynamic and sparse operator for vision applications\.InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,pp\. 5652–5661\.Cited by:[Occupancy Prediction Methods](https://arxiv.org/html/2609.21470#Sx2.SSx1.p1.1)\.
- Xuet al\.\(2025\)T\. Xu, H\. Lu, X\. Yan, Y\. Cai, B\. Liu, and Y\. ChenOcc\-llm: enhancing autonomous driving with occupancy\-based large language models\.arXiv preprint arXiv:2502\.06419\.Cited by:[Occupancy Prediction Methods](https://arxiv.org/html/2609.21470#Sx2.SSx1.p1.1)\.
- Yeet al\.\(2024\)Z\. Ye, T\. Jiang, C\. Xu, Y\. Li, and H\. ZhaoCVT\-occ: cost volume temporal fusion for 3d occupancy prediction\.External Links:2409\.13430,[Link](https://arxiv.org/abs/2409.13430)Cited by:[Table 1](https://arxiv.org/html/2609.21470#Sx4.T1.1.11.1.1),[Benchmarking on RiskOcc4D\-nuScenes](https://arxiv.org/html/2609.21470#Sx5.SSx2.p1.1)\.
- Yuet al\.\(2023\)Z\. Yu, C\. Shu, J\. Deng, K\. Lu, Z\. Liu, J\. Yu, D\. Yang, H\. Li, and Y\. ChenFlashocc: fast and memory\-efficient occupancy prediction via channel\-to\-height plugin\.arXiv preprint arXiv:2311\.12058\.Cited by:[Occupancy Prediction Methods](https://arxiv.org/html/2609.21470#Sx2.SSx1.p1.1),[Table 1](https://arxiv.org/html/2609.21470#Sx4.T1.1.12.1),[Table 1](https://arxiv.org/html/2609.21470#Sx4.T1.1.9.1.1),[Benchmarking on RiskOcc4D\-nuScenes](https://arxiv.org/html/2609.21470#Sx5.SSx2.p1.1),[Benchmarking on RiskOcc4D\-nuScenes](https://arxiv.org/html/2609.21470#Sx5.SSx2.p4.1)\.
- Zhanget al\.\(2023a\)C\. Zhang, J\. Yan, Y\. Wei, J\. Li, L\. Liu, Y\. Tang, Y\. Duan, and J\. LuOccnerf: advancing 3d occupancy prediction in lidar\-free environments\.arXiv preprint arXiv:2312\.09243\.Cited by:[Occupancy Prediction Methods](https://arxiv.org/html/2609.21470#Sx2.SSx1.p1.1)\.
- Zhanget al\.\(2023b\)Y\. Zhang, Z\. Zhu, and D\. DuOccformer: dual\-path transformer for vision\-based 3d semantic occupancy prediction\.InProceedings of the IEEE/CVF International Conference on Computer Vision,pp\. 9433–9443\.Cited by:[Occupancy Prediction Methods](https://arxiv.org/html/2609.21470#Sx2.SSx1.p1.1)\.
- Zhaoet al\.\(2024\)Y\. Zhao, P\. Zheng, and Z\. Gong3D occupancy and flow prediction based on forward view transformation\.InCVPR 2024 Workshop on Foundation Models for Autonomous Systems,Cited by:[Occupancy Prediction Methods](https://arxiv.org/html/2609.21470#Sx2.SSx1.p1.1)\.
- Zhenget al\.\(2025\)W\. Zheng, W\. Chen, Y\. Huang, B\. Zhang, Y\. Duan, and J\. LuOccworld: learning a 3d occupancy world model for autonomous driving\.InEuropean Conference on Computer Vision,pp\. 55–72\.Cited by:[Occupancy Prediction Methods](https://arxiv.org/html/2609.21470#Sx2.SSx1.p1.1)\.
- Zhenget al\.\(2024\)W\. Zheng, R\. Song, X\. Guo, C\. Zhang, and L\. ChenGenAD: generative end\-to\-end autonomous driving\.External Links:2402\.11502,[Link](https://arxiv.org/abs/2402.11502)Cited by:[ROIDrive Framework](https://arxiv.org/html/2609.21470#Sx3.SSx1.p1.1),[ROIDrive Framework](https://arxiv.org/html/2609.21470#Sx3.SSx1.p4.1),[Experimental Design](https://arxiv.org/html/2609.21470#Sx5.SSx3.SSSx1.p1.1),[Table 2](https://arxiv.org/html/2609.21470#Sx5.T2),[Conclusion](https://arxiv.org/html/2609.21470#Sx6.p1.1)\.
- Zouet al\.\(2025\)J\. Zou, S\. Chen, B\. Liao, Z\. Zheng, Y\. Song, L\. Zhang, Q\. Zhang, W\. Liu, and X\. WangDiffusionDriveV2: reinforcement learning\-constrained truncated diffusion modeling in end\-to\-end autonomous driving\.arXiv preprint arXiv:2512\.07745\.External Links:[Document](https://dx.doi.org/10.48550/arXiv.2512.07745),[Link](https://arxiv.org/abs/2512.07745)Cited by:[Auxiliary Perception Supervision in End\-to\-End Driving Models](https://arxiv.org/html/2609.21470#Sx2.SSx2.p1.1)\.