🚗 Autonomous Driving¶
🧠 NeurIPS2026 · 52 paper notes
📌 Same area in other venues: 🎞️ ECCV2026 (161) · 📷 CVPR2026 (176) · 🔬 ICLR2026 (54) · 🧪 ICML2026 (8) · 🤖 AAAI2026 (63) · 🧠 NeurIPS2025 (47)
🔥 Top topics: Autonomous Driving ×18 · Multimodal/VLM ×6 · Agents ×4 · Reasoning ×4 · Diffusion Models ×2
- AutoExpert: Automating 3D LiDAR Annotation from Expert-Crafted Guidelines
-
AutoExpert turns textual expert rules and a few 2D examples into an adapted detector, then uses a vision-language model to supply instance-specific size and orientation priors for fixed-size multi-hypothesis search, achieving 25.4 mAP3D on AutoExpert-nuScenes without target-task 3D training annotations.
- Autonomous Driving Research Requires a Community-Driven Data Paradigm
-
This position paper systematically analyzes the structural disconnect between citation-monopolizing commercial benchmarks and over 640 unindexed long-tail academic driving datasets, advocating for a community-driven collaborative data ecosystem that decouples raw data collection from task annotation with standardized interoperability and shared infrastructure.
- Beyond Chamfer Distance: Granular Order-aware Evaluation Metric For Online Mapping
-
To address the critical flaws of Chamfer distance (CD) and mAP—namely lack of point-order awareness, susceptibility to outliers, and hard thresholding that masks localization errors—this paper introduces SOSPA, a metric-compliant order-aware single-instance distance, and PLD, a soft multi-instance metric that decomposes into interpretable localization and detection error terms.
- Beyond Chamfer Distance: Granular Order-Aware Evaluation Metric for Online HD Mapping
-
Addressing the lack of order awareness in Chamfer Distance and the loss of positional fidelity caused by hard thresholding in AP, this paper introduces SOSPA, a metric-axiom-compliant ordered polyline distance, and PLD, a multi-instance probabilistic soft-matching metric that decouples mapping error into localization and detection components.
- Beyond Self-Play and Scale: A Behavior Benchmark for Generalization in Autonomous Driving
-
Addressing the disconnection between at-scale RL driving policies and standardized benchmarks, as well as the simplistic traffic behaviors of current evaluations, this paper introduces BehaviorBench—featuring a nuPlan evaluation interface, an interaction-rich WOMD split (Interactive1k), and heterogeneous traffic agents—uncovering self-play RL generalization failures and establishing a robust hybrid PDM+PPO planner.
- BitMTP: When Multi-Token Prediction Meets Low-Bit Large Language Models
-
Addressing the quality degradation and runtime overheads when naively combining multi-token prediction with low-bit backbones, BitMTP introduces an algorithm–system co-design framework featuring joint NTP-MTP training, flexible decoding strategies, and an accepted-prefix-aligned runtime that delivers 2.1× to 2.4× decode speedups at comparable quality.
- Bridging Scene Generation and Planning: Driving with World Model via Unifying Visual Geometry
-
WorldDrive bridges scene generation and real-time planning by coupling visual and motion representations in a trajectory-aware driving world model, inheriting these mature representations to a downstream planner, and employing a future-aware rewarder to enable millisecond-level foresight evaluation without diffusion denoising.
- CausalDriveBench: Evaluating Causal Reasoning in Vision-Language-Action Models for Autonomous Driving
-
The first benchmark to evaluate causal reasoning in autonomous driving vision-language-action (VLA) models grounded in Pearl's Causal Hierarchy (PCH), featuring causal scene graphs with active/dormant/distractor partitions, 7,285 verified QA pairs, and 1,000 audited counterfactual trajectories that expose severe decoupling between language rationales and motor planning.
- Controllable Road Marking Generation
-
This paper introduces the controllable road marking generation task, combining a latent rectified-flow DiT with topology-aware soft-skeleton supervision and training-free Structured Gaussian Render (SGR) to synthesize high-fidelity lane dividers, road dividers, and pedestrian crossings from drivable geometry, outer-ring context, and text descriptions.
- CoWorld-VLA: Thinking in a Multi-Expert World Model for Autonomous Driving
-
CoWorld-VLA formulates intermediate world reasoning as a multi-expert Latent CoT spanning semantic, geometric, dynamic, and trajectory priors within a VLM, and leverages a two-stream hierarchical diffusion planner to directly condition continuous trajectory generation on these physical world representations.
- CRAFT: Counterfactual-to-Interactive Reinforcement Fine-Tuning for Driving Policies
-
Addressing the severe distribution shift encountered when deploying open-loop imitation driving policies into closed-loop interactive environments, CRAFT introduces a proxy-residual reinforcement fine-tuning framework that unifies dense group-normalized counterfactual advantages with grounded, dual-clipped residual corrections from safety-critical closed-loop events, achieving state-of-the-art gains across diverse policy architectures on Bench2Drive.
- CRISP: Fixing Flying Pixels in Latent LiDAR Generation via Diffusion Decoding
-
Addressing the severe "flying pixel" artifacts caused by convolutional VAE decoders oversmoothing radial depth boundaries in latent LiDAR pipelines, CRISP introduces a backbone-agnostic plug-in diffusion decoder that drastically sharpens boundary geometry and cuts perceptual error by over 50% while keeping the encoder and latent generator completely frozen.
- Drift-Resistant Navigation World Model with Anchored Epipolar Guidance
-
Addressing error accumulation (perceptual drift) and viewpoint misalignment (geometric drift) in autoregressive navigation world models, DR-NWM proposes sparse anchor-guided chunked interpolation combined with bidirectional epipolar line intersection attention masks, enabling high-fidelity, geometry-grounded long-horizon future prediction and downstream planning gains without explicit 3D supervision.
- DriveDreamer-Policy: A Geometry-Grounded World–Action Model for Unified Generation and Planning
-
DriveDreamer-Policy unifies multimodal LLM reasoning with lightweight generative experts by introducing explicit monocular depth generation as a 3D geometric scaffold, orchestrating a causal "depth → video → action" information flow to achieve high-fidelity future imagination and robust end-to-end motion planning.
- DriveFuture: Future-Aware Latent World Models for Autonomous Driving
-
DriveFuture addresses the fundamental limitation in autonomous driving where latent world models treat future states merely as prediction targets rather than planning conditions, proposing to condition diffusion trajectory generation directly on future world states alongside LatentAlign training and Progressive Foresight Guidance (PFG) to break inference circularity, achieving 55.5 EPDMS on NAVSIM-v2 navhard.
- DriveHierarchy: A Benchmark for Diagnosing VLM Driving Capabilities from Open-Loop Understanding to Closed-Loop Execution
-
DriveHierarchy organizes vision-language driving capabilities into perceptual grounding, contextual memory, mental reasoning, and closed-loop execution, diagnoses 15 models through unified open-loop tasks and 100 interactive simulation scenarios, and shows on Qwen3-VL-8B that repairing selected open-loop weaknesses can raise the closed-loop composite score from 0.902 to 8.021, without establishing real-world driving safety.
- DriveMind: Mind-Evolving Belief Tracking for Closed-Loop Autonomous Driving
-
DriveMind borrows principles from particle filtering to maintain a time-evolving belief state over multiple natural-language hypotheses, decoupling high-level semantic reasoning from high-frequency diffusion trajectory generation through an asynchronous fast-slow VLA architecture to resolve premature commitment in ambiguous closed-loop driving.
- DriveSpatial: A Benchmark for Spatiotemporal Intelligence in VLMs for Autonomous Driving
-
DriveSpatial establishes the first comprehensive benchmark assessing the spatiotemporal intelligence of vision-language models in multi-view autonomous driving scenes across five public datasets, revealing a staggering 28.4% performance deficit behind human capability and pinpointing internal cognitive scene construction as the primary upstream bottleneck.
- Driving on Memory
-
This paper introduces MemoryDrivoR, an analytic probe model for auditing end-to-end autonomous driving benchmarks that strips online camera inputs and relies solely on pose-retrieved latent memories of past traversals plus initial ego status, exposing critical vulnerabilities where models near SOTA performance on NAVSIM without observing the evaluated scene.
- Driving Video Retrieval for Complex Queries with Structured Grounding
-
STRIVE-D calibrates motion rules against a weak-label ranking objective on separate driving videos, reuses or adapts those rules per query, and fuses them with visual and lexical retrieval, raising DrivingDojo Acc@1 from the main table's strongest dense baseline of 14.6% to 26.8%; it does not simply ask an LLM to inspect videos and determine true geometry.
- FlashBoB: I/O-Efficient Exact Backward-over-Backward for Softmax Attention
-
FlashBoB addresses memory explosion and global dependencies in softmax attention second-order back-propagation (backward-over-backward) by exploiting a hierarchical affine structure to collapse the computation into an exact two-pass tiled schedule, eliminating all \(N \times N\) intermediates and matching the large-cache I/O lower bound.
- From Representational Complementarity to Dual Systems: Synergizing VLM and Vision-Only Backbones for End-to-End Driving
-
The paper uncovers that downstream policy learning expands a shared decision subspace between VLMs and vision-only encoders while preserving functionally complementary residual factors, and exploits this via trajectory scoring (HybridDriveVLA) and fast-slow adaptive scheduling (DualDriveVLA).
- Generative Scenario Rollouts for End-to-End Autonomous Driving
-
GeRo is a plug-and-play Vision-Language-Action (VLA) framework that jointly unifies motion planning with language-grounded autoregressive scenario generation in a compact latent token space, utilizing temporal consistency regularization and multi-objective GRPO reinforcement learning to dramatically improve long-horizon planning robustness and language-action alignment.
- Grounding Driving VLA via Inverse Kinematics
-
By treating driving trajectory generation as an inverse kinematics problem requiring both initial and terminal visual boundary conditions, this work introduces a next visual state prediction objective and a visual-only bottleneck conditional diffusion network, enabling a 0.5B VLA model to achieve driving performance comparable to 7B–8B baselines on NAVSIM and nuScenes.
- Intend, Reflect, Refine: An Adaptive Multimodal Reflection Framework for Autonomous Driving
-
IRR-Drive proposes the first adaptive multimodal reflection framework for end-to-end autonomous driving that combines textual reasoning with discrete semantic BEV forecasting, achieving grounded trajectory refinement and optimal efficiency-performance balance via difficulty-aware GSPO reinforcement learning to set new state-of-the-art results on NAVSIM v1 and v2.
- LaST-VLA: Thinking in Latent Spatio-Temporal Space for Vision-Language-Action in Autonomous Driving
-
LaST-VLA shifts autonomous driving reasoning from discrete textual chains to a continuous latent spatio-temporal space, distilling 3D geometric and world model dynamic priors via dual adapters coupled with progressive two-stage SFT and GRPO, achieving new state-of-the-art performance on NAVSIM benchmarks.
- Learning to Drive in New Cities Without Human Demonstrations
-
NOMAD adapts driving policies to unseen target cities without collecting target-city human demonstrations by combining map-conditioned scenario generation with KL-regularized multi-agent self-play, expanding the success-realism Pareto frontier beyond zero-shot and imitation baselines.
- LiFi: LiDAR Generation from Multi-View Images via Geometric and Semantic Collaborative Guidance
-
To tackle cross-modal geometric alignment and perspective distortion when synthesizing LiDAR point clouds from multi-view camera images, LiFi introduces a decoupled latent diffusion framework featuring range-view-centered reverse depth sampling, uncertainty-modulated geometric encoding, cross-view semantic interaction, and balanced classifier-free guidance, establishing high point-level geometric fidelity and cross-modal consistency.
- LogSTOP: Temporal Scores over Prediction Sequences for Matching and Retrieval
-
Addressing the exponential verification complexity and extreme noise sensitivity of classical temporal logic model checking, LogSTOP recursively computes sequence-level temporal property scores via log-probability semantics in \(\mathcal{O}(T \cdot |\varphi|)\) linear time, leveraging adaptive thresholds and subsequence max-pooling to outperform large multimodal models and formal verification baselines in video and audio temporal matching and retrieval.
LR-V2X: Loss-Resilient Collaborative Perception under Low-Bandwidth Communication
- MAPLE: Latent Multi-Agent Play for End-to-End Autonomous Driving
-
MAPLE introduces a simulator-free, multi-agent closed-loop training framework operating entirely within the latent token space of a vision-language-action (VLA) model, achieving new state-of-the-art closed-loop driving performance on Bench2Drive via diversity-regularized group relative policy optimization (GRPO).
- Markovian Dynamics Enforcer: Feasibility Preserving Correction on Learned Dynamics Manifolds
-
MaDE learns inverse dynamics and physical residuals from state transitions without control labels, then corrects predicted trajectories in control space so that outputs satisfy its learned discrete dynamics map by construction while reducing inequality violations on a best-effort basis; known-model residuals fall substantially on inD, but position errors increase, without establishing true-physics correctness or safety.
- MCSplat: Multi-View Photometric and Geometric Consistent Feed-Forward Gaussian Splatting for Driving Scenes
-
Addressing cross-camera photometric inconsistency and unstable local geometry in feed-forward 3D Gaussian splatting for surround-view driving scenes, MCSplat introduces depth-guided canonical appearance construction, multi-scale Bilateral Grid Transformer appearance recovery, and Fisher-information-guided geometric self-supervision, achieving state-of-the-art novel view synthesis and cross-view geometric consistency on Waymo and nuScenes.
- MESSENGER: Memory-Enhanced Sequential Scene Flow Estimation via Autoregressive Next-Frame ForecastinG
-
MESSENGER maintains a coarse-resolution temporal memory buffer storing historical flow vectors and hidden states, retrieves next-frame motion priors via autoregressive cross-attention, and deploys discrepancy-driven uncertainty reweighting to suppress compounding drift, achieving constant GPU memory footprint and state-of-the-art long-horizon future motion extrapolation.
- Motion Cues from Image-based Point Tracking for LiDAR Scene Flow Estimation
-
Addressing the vulnerability of ray-casting dynamic mapping to LiDAR sparsity and occlusions in self-supervised scene flow estimation, TrackCue extracts dense image-space trajectories via video point tracking, isolates true motion from ego-induced parallax through rigid trajectory compensation, and lifts validated cues back to 3D point clouds to provide high-fidelity supervisory pseudo-labels.
- nuReasoning: A Reasoning-Centric Dataset and Benchmark for Long-Tail Autonomous Driving
-
To resolve the vulnerability of end-to-end autonomous driving in rare and safety-critical cases, this paper presents nuReasoning, the first large-scale benchmark of 20,000 real-world long-tail driving clips with dense spatial-decision-counterfactual reasoning annotations, demonstrating that reasoning supervision significantly enhances downstream VLA trajectory planning safety.
- OmniTraffic: A Controllable Generation Pipeline and Benchmark for Spatio-Temporal Traffic Reasoning
-
OmniTraffic reconstructs 12 real-world intersections into editable 3D simulation environments complemented by real surveillance footage, establishing an 8-million sample VQA generation pipeline and a 3K human-verified benchmark spanning perception, spatio-temporal reasoning, and decision support, revealing substantial human-model capability gaps in topology grounding and temporal dynamics.
- One Step is Enough: Multi-Agent Reinforcement Learning Based on One-Step Policy Optimization for Order Dispatch on Ride-Sharing Platforms
-
Addressing the severe bias and instability of neural critic networks in autonomous vehicle order dispatch, this paper establishes the value homogeneity of homogeneous fleets under rapid transportation network mixing and proposes critic-free Single-Trajectory GRPO (ST-GRPO) along with One-Step Policy Optimization (OSPO), achieving SOTA service rates and pickup efficiency with minimal GPU footprint.
- One World, Dual Timeline: Decoupled Spatio-Temporal Gaussian Scene Graph for 4D Cooperative Driving Reconstruction
-
To tackle temporal asynchrony between independent hardware clocks of vehicle and infrastructure cameras in cooperative driving, DUST introduces a decoupled spatio-temporal Gaussian scene graph that pairs shared canonical 3D Gaussians with source-specific dual pose timelines, eliminating gradient conflicts and dynamic ghosting.
- PointForward: Feedforward Driving Reconstruction through Point-Aligned Representations
-
To tackle multi-view geometric conflicts, layering artifacts, and fragmented dynamic motions in per-pixel feedforward 3D Gaussian Splatting for driving scenes, PointForward introduces sparse 3D point-aligned queries initialized in world space, incorporates 3D bounding box scene graphs for canonical dynamic modeling, and fuses multi-view spatial-temporal cues to achieve state-of-the-art single-pass reconstruction.
- Progressive Risk Estimation for Accident Anticipation
-
PRE-ACT uses continuous distance-to-event progress supervision and within-video clip ranking to improve traffic risk prediction from past-only sliding windows, raising CAP mAUC from TOP's 0.429 to 0.481 under FPR ≤ 0.1 and introducing a Separation Score to examine false-alarm tendencies across complete risk curves.
- ReflectDrive-2: Reinforcement-Learning-Aligned Self-Editing for Discrete Diffusion Driving
-
ReflectDrive-2 introduces a masked discrete diffusion planner for autonomous driving that represents trajectories as discrete tokens and performs in-place self-editing (AutoEdit) without auxiliary networks, using full-rollout reinforcement learning to amplify the post-editing gain from +0.3 to +1.9 PDMS and reaching 91.0 PDMS on NAVSIM with camera-only input at 30.2 ms latency on NVIDIA Thor.
- RESBev: Making BEV Perception More Robust
-
RESBev introduces a plug-and-play temporal framework that enhances BEV perception robustness against diverse natural corruptions and adversarial perturbations by forecasting a clean semantic prior in a compact latent space and selectively recovering valid scene context via query-guided cross-attention.
- Robust Residual Correction via Selective Deployment for Time Series Forecasting
-
To overcome the "Corrector's Dilemma" where always-on residual updates overfit noise and degrade forecasts, this paper couples a direction-aware hybrid corrector with a validation-calibrated selective deployment firewall, achieving substantial MSE reductions and high non-degradation rates across diverse benchmarks.
- RVLoss: Runoff Vote Loss for Self-Supervised LiDAR Scene Flow Estimation
-
Addressing the breakdown of motion rigidity in self-supervised LiDAR scene flow caused by nearest neighbor search, RVLoss introduces a two-stage runoff voting mechanism that generates cluster-wise rigid and continuous free-form flow pseudo-labels, improving self-supervised feedforward scene flow accuracy by up to 20% without requiring architectural modifications or handcrafted regularizers.
- Scaling Curriculum Learning For Autonomous Driving
-
CL4AD integrates unsupervised environment design into batched GPU-accelerated self-play simulators for autonomous driving, introducing collision-free Learnability-Hard and behavioral realism utilities to prioritize learning frontiers and accelerating RL training to 99% success by up to 77% over domain randomization.
- Scene-Adaptive VLA: Efficient Autonomous Driving via Dynamic Layer Routing
-
To alleviate the severe onboard computational burden of Vision-Language-Action (VLA) models in autonomous driving, Scene-Adaptive VLA introduces dynamic layer routing conditioned on temporal visual variations and inter-frame decision deviations, slashing inference computational overhead by 53.5% on Bench2Drive with negligible impact on driving and linguistic capabilities.
- SOAR: Regression-based LiDAR Relocalization for UAVs
-
Addressing the failure of existing ground vehicle LiDAR relocalization methods under arbitrary yaw rotations, drastic altitude shifts, and irregular trajectories in UAV flights, SOAR introduces a map-free scene coordinate regression framework featuring coordinate-independent serialization and locality-preserving sliding window attention, achieving meter-level localization accuracy on realistic UAV benchmarks without explicit 3D maps.
- The Road Ahead in Autonomous Driving: The KITScenes Multimodal Dataset
-
KITScenes Multimodal is a high-fidelity European autonomous driving dataset featuring a synchronized 72.5 MP global-shutter camera suite, 400 m long-range lidars, and 4D imaging radars, paired with production-grade Lanelet2 HD maps covering 62 km² with pixel-accurate 3D regulatory elements and four deployment-oriented spatial benchmarks.
- TopoCurve: Geometry-Aware Topology Reasoning via Bézier Curves in Autonomous Driving
-
Addressing the issues of jagged geometry, missing analytical tangents, and sparse topology supervision inherent in polyline representations, TopoCurve reparameterizes lanes as endpoint-fixed cubic Bézier curves, integrating Fourier-encoded geometric bias, curve-distributed deformable attention, and redundant supervision to establish a new state-of-the-art of 50.6 OLS on OpenLane-V2 without any post-processing.
- Towards Effective and Transferable Physical Camouflage against Multi-View BEV-based 3D Perception in Autonomous Driving
-
BEVCA presents an adversarial camouflage framework against multi-view BEV-based 3D perception by uniting cross-coordinate multi-view neural rendering with a polar mask-guided BEV feature suppression loss, producing physical textures that transfer effectively across victim models and perception tasks.
- UNIVERSE: Unified Video Action Models for Autonomous Driving with Flexible Mask-Modulated Modality Generation
-
UNIVERSE proposes a unified single-DiT video-action model that co-trains future video latents and ego-trajectory tokens within shared generative parameters, leveraging a modality-decoupling visibility mask to prevent causal leakage and enable a 4.3x speedup via trajectory-only inference.