🚗 Autonomous Driving¶

🧪 ICML2026 · 8 paper notes

📌 Same area in other venues: 📷 CVPR2026 (157) · 🔬 ICLR2026 (50) · 🤖 AAAI2026 (56) · 🧠 NeurIPS2025 (47) · 📹 ICCV2025 (91) · 🧪 ICML2025 (10)

🔥 Top topics: Autonomous Driving ×3 · Reinforcement Learning ×2 · Navigation ×2

CoIRL-AD: Collaborative-Competitive Imitation-Reinforcement Learning in Latent World Models for Autonomous Driving: CoIRL-AD utilizes two independent actors to handle Imitation Learning (IL) and Reinforcement Learning (RL) respectively, relying on a latent world model to "imagine" future trajectories for calculating long-range rewards for RL. A "leader-follower" competitive mechanism allows both actors to transfer beneficial behaviors to each other. This approach successfully integrates RL into end-to-end driving using offline real-world driving data without an external simulator, achieving significant improvements in cross-city generalization and long-tail scenarios.
Constrained Multi-Objective Reinforcement Learning with Max-Min Criterion: This paper unifies "max-min multi-objective fairness" and "hard constraint satisfaction" into a single MORL framework. By reformulating the problem as a convex program via occupancy measures, the authors derive a dual convex optimization problem over weights \((u,w)\). This allows a projected gradient descent algorithm to simultaneously achieve fairness and constraint feasibility with theoretical guarantees of geometric convergence.
DeepSight: Long-Horizon World Modeling via Latent States Prediction for End-to-End Autonomous Driving: DeepSight shifts "future world prediction" from explicit pixel reconstruction (single-frame codebook) to parallel implicit multi-frame prediction of DINOv3 semantic features in BEV space. Combined with an on-demand Adaptive Chain-of-Thought, it achieves a Driving Score of 86.23 (+7.39) and a Success Rate of 71.36% (+13.63) on the Bench2Drive closed-loop benchmark while adding only ~4% inference latency.
Mitigating Error Accumulation in Continuous Navigation via Memory-Augmented Kalman Filtering: This work reformulates step-by-step prediction in continuous UAV VLN as a closed-loop "recursive Bayesian estimation = GRU prior + memory likelihood + learnable Kalman gain." By fine-tuning on only 10% of the data in TravelUAV, the Success Rate (SR) of L1-Full is improved from 17.6% to 25.9%, while the positional drift—which typically accumulates continuously after 100 steps—is flattened to 30–40 meters.
Plug-and-Play Label Map Diffusion for Universal Goal-Oriented Navigation: This paper proposes PLMD: a framework that merges BEV semantic and obstacle maps into a unified Label Map. It utilizes DDPM, modulated by obstacle priors, to complete semantic and obstacle labels in unexplored regions. As a plug-and-play module, it can be integrated with any GON policy and consistently achieves new SOTA results on HM3D/MP3D across three tasks: ON, IIN, and MRON.
RoCA: Robust Cross-Domain End-to-End Autonomous Driving: RoCA attaches a plug-and-play module based on Gaussian Processes (GP) to end-to-end autonomous driving models. By learning a set of basis tokens and corresponding trajectories that cover diverse scenarios, it probabilistically infers future trajectories based on similarity for new scenarios. This approach uses GP uncertainty for regularization to enhance generalization during source domain training and enables efficient adaptation via pseudo-labels and active learning in new domains, without requiring LLMs or increasing inference overhead.
Threshold-Based Exclusive Batching for LLM Inference: This paper systematically characterizes the performance crossover conditions between mixed batching (MB) and exclusive batching (EB) in LLM inference. It proves that on bandwidth-constrained GPUs, co-batching prefill and decode stages slows down Attention due to bandwidth contention. Consequently, the authors derive an optimal phase-switching threshold \(\theta^*\) and a memory-safe batch size based on the hazard rate, designing an online adaptive scheduler EB+. This scheduler improves throughput by up to 41.9% on bandwidth-constrained hardware and up to 36.4% under non-stationary traffic compared to MB.
TSRBench: A Comprehensive Multi-task Multi-modal Time Series Reasoning Benchmark for Generalist Models: TSRBench constructs a time series reasoning benchmark covering 14 domains, 4 major dimensions (Perception/Reasoning/Prediction/Decision-making), 15 tasks, and 4,125 questions. It supports four input modalities (Text, Visual, Text+Image, Embedding) and systematically evaluates 30+ mainstream LLMs, VLMs, and TSLLMs. It reveals that "scaling holds in perception/reasoning but fails in prediction" and that "text and visual modalities are highly complementary, yet current models struggle to fuse them."