Technical deep dives into the evolution of autonomous driving from modular pipelines to end-to-end systems, VLA architectures, and generative planning.
Foundational Arguments
| Article | Core Thesis |
|---|---|
| Why Generative Planning? | The feasible set is non-convex; regression fundamentally fails |
| Trajectory Tokenization for AR Planning | Clustering, matching, and the AR+Diffusion paradigm |
| RL Policy Optimization for E2E | From REINFORCE to GRPO for driving |
| E2E Architecture Evolution | V2.0 decoder selection to V3.0 VLA integration |
Model Architecture & Conditioning
| Article | Topic |
|---|---|
| Alpamayo VLA | Vision-Language-Action for driving |
| RL: DPO to Self-Improvement | Post-training pipeline for driving |
| Condition Consumption in Planning | From timestep τ to navigation injection |
| VLM Temporal Memory | Temporal memory mechanisms for VLMs |
End-to-End Driving
| Article | Topic |
|---|---|
| ReflectDrive-2 | Discrete diffusion for end-to-end driving (Li Auto) |
Training & Scaling
| Article | Topic |
|---|---|
| VLA Train Recipe at Billion Scale | Mapping TRI co-training findings (arXiv:2602.01067) + Qwen-VLA recipe to driving VLA at billion-sample scale |
Generative Planning
| Article | Topic |
|---|---|
| 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 |