Technical deep dives into the evolution of autonomous driving from modular pipelines to end-to-end systems, VLA architectures, and generative planning.

Foundational Arguments

ArticleCore Thesis
Why Generative Planning?The feasible set is non-convex; regression fundamentally fails
Trajectory Tokenization for AR PlanningClustering, matching, and the AR+Diffusion paradigm
RL Policy Optimization for E2EFrom REINFORCE to GRPO for driving
E2E Architecture EvolutionV2.0 decoder selection to V3.0 VLA integration

Model Architecture & Conditioning

ArticleTopic
Alpamayo VLAVision-Language-Action for driving
RL: DPO to Self-ImprovementPost-training pipeline for driving
Condition Consumption in PlanningFrom timestep τ to navigation injection
VLM Temporal MemoryTemporal memory mechanisms for VLMs

End-to-End Driving

ArticleTopic
ReflectDrive-2Discrete diffusion for end-to-end driving (Li Auto)

Training & Scaling

ArticleTopic
VLA Train Recipe at Billion ScaleMapping TRI co-training findings (arXiv:2602.01067) + Qwen-VLA recipe to driving VLA at billion-sample scale

Generative Planning

ArticleTopic
扩散模型与自动驾驶规划From denoising mathematics to trajectory generation
Flow Matching 训练的三轴分解Cone 几何下三轴正交分解(数据广度深度 / schedule / 物理 horizon),endBend 是 depth 在三轴乘性 dilute 后的可观察残余