自动驾驶 on 朝花夕拾

自动驾驶 on 朝花夕拾 https://xwlu.github.io/tags/%E8%87%AA%E5%8A%A8%E9%A9%BE%E9%A9%B6/ Recent content in 自动驾驶 on 朝花夕拾 Hugo en-us Fri, 15 May 2026 14:46:23 +0800 RAD-2: Scaling Reinforcement Learning in a Generator-Discriminator Framework https://xwlu.github.io/post/robotics/e2e/rad_2/ Fri, 24 Apr 2026 00:00:00 +0000 https://xwlu.github.io/post/robotics/e2e/rad_2/ 想象一下，你在教一个新手司机开车。用模仿学习（Imitation Learning）的方式，就像给他一大堆"老司机录像"反复看——他能学到很多驾驶技巧，但问题是，<strong>他只知道"该怎么开"，却从来没有为自己的失误付出过代价</strong>。 MOSAIC: 基于规模感知的自动驾驶数据挑选魔法阵 https://xwlu.github.io/post/robotics/e2e/mosaic/ Tue, 21 Apr 2026 00:00:00 +0000 https://xwlu.github.io/post/robotics/e2e/mosaic/ 在物理人工智能（Physical AI）领域，特别是自动驾驶，业界一直以来的做法就是 <strong>“疯狂喂数据”</strong>。但这就带来了一个巨大的痛点：<strong>“偏科"与"效能黑盒”</strong>。 DriveTransformer: Unified Transformer for Scalable End-to-End Autonomous Driving https://xwlu.github.io/post/robotics/e2e/drive_transformer/ Mon, 20 Apr 2026 00:00:00 +0000 https://xwlu.github.io/post/robotics/e2e/drive_transformer/ 想象一下，传统的自动驾驶系统就像一个<strong>刻板的流水线工厂</strong>： SparseDriveV2: Scoring is All You Need for End-to-End Autonomous Driving https://xwlu.github.io/post/robotics/e2e/sparse_drive_v2/ Mon, 20 Apr 2026 00:00:00 +0000 https://xwlu.github.io/post/robotics/e2e/sparse_drive_v2/ 端到端自动驾驶的多模态规划中，江湖上原本分为两派： DiffusionDriveV2: Truncated Diffusion Model for End-to-End Autonomous Driving https://xwlu.github.io/post/robotics/e2e/diffusion_drive_v2/ Tue, 17 Mar 2026 00:00:00 +0000 https://xwlu.github.io/post/robotics/e2e/diffusion_drive_v2/ 自动驾驶规划面临一个经典的两难困境：<strong>多样性 vs 质量</strong> End-to-End Autonomous Driving without Costly Modularization and 3D Manual Annotation https://xwlu.github.io/post/robotics/e2e/uad/ Tue, 17 Mar 2026 00:00:00 +0000 https://xwlu.github.io/post/robotics/e2e/uad/ 在聊 UAD 之前，咱们先看看现有的端到端自动驾驶老大哥们（比如 UniAD）。虽然它们号称"端到端"，但骨子里还是在模仿传统流水线，设计了层层递进的 <strong>感知 → 预测 → 规划</strong> 子任务。 Epona: Autoregressive Diffusion World Model for End-to-End Autonomous Driving https://xwlu.github.io/post/robotics/e2e/epona/ Tue, 17 Mar 2026 00:00:00 +0000 https://xwlu.github.io/post/robotics/e2e/epona/ <blockquote> <p>端到端自动驾驶的视频生成与轨迹规划</p></blockquote> FutureX: Enhance End-to-End Autonomous Driving via Latent Chain-of-Thought World Model https://xwlu.github.io/post/robotics/e2e/future_x/ Tue, 17 Mar 2026 00:00:00 +0000 https://xwlu.github.io/post/robotics/e2e/future_x/ 各位乘客请系好安全带！今天我们要深入了解一款拥有"老司机思维"的自动驾驶大模型——<strong>FutureX</strong>。 HiP-AD: Hierarchical and Multi-granularity Planning with Deformable Attention https://xwlu.github.io/post/robotics/e2e/hi_p_ad/ Tue, 17 Mar 2026 00:00:00 +0000 https://xwlu.github.io/post/robotics/e2e/hi_p_ad/ 现在的端到端自动驾驶（E2E-AD）界有个普遍的怪现象：<strong>“应试教育"满分，“实战上路"拉胯。</strong> MomAD: Momentum-Aware Planning in End-to-End Autonomous Driving https://xwlu.github.io/post/robotics/e2e/mom_ad/ Tue, 17 Mar 2026 00:00:00 +0000 https://xwlu.github.io/post/robotics/e2e/mom_ad/ <blockquote> <p>论文标题：《Don’t Shake the Wheel: Momentum-Aware Planning in End-to-End Autonomous Driving》</p></blockquote> MotionLM: Multi-Agent Motion Forecasting as Language Modeling https://xwlu.github.io/post/robotics/e2e/motion_lm/ Tue, 17 Mar 2026 00:00:00 +0000 https://xwlu.github.io/post/robotics/e2e/motion_lm/ 把自动驾驶的多智能体轨迹预测，变成一场"文字接龙"游戏——用语言模型预测下一个动作词的方式，来预测车辆和行人的未来轨迹。 ResAD: Normalized Residual Trajectory Modeling for End-to-End Autonomous Driving https://xwlu.github.io/post/robotics/e2e/res_ad/ Tue, 17 Mar 2026 00:00:00 +0000 https://xwlu.github.io/post/robotics/e2e/res_ad/ 目前的端到端自动驾驶模型（E2EAD）大多试图回答同一个问题："<strong>未来的轨迹是什么？</strong>" 它们直接从传感器数据预测车辆未来几秒钟的绝对坐标点 $(x, y)$。 World4Drive - 无需感知标注的端到端自动驾驶世界模型 https://xwlu.github.io/post/robotics/e2e/world4_drive/ Tue, 17 Mar 2026 00:00:00 +0000 https://xwlu.github.io/post/robotics/e2e/world4_drive/ <blockquote> <p>这篇论文的核心思想可以概括为：<strong>如何培养一个会自己"脑补"未来、且极具空间方向感的老司机</strong>。</p></blockquote> LAW - Enhancing End-to-End Autonomous Driving with Latent World Model https://xwlu.github.io/post/robotics/e2e/law/ Mon, 01 Jan 0001 00:00:00 +0000 https://xwlu.github.io/post/robotics/e2e/law/ 通过在潜空间进行动作引导的未来特征预测，实现了无需标注的深度场景特征学习，显著提升了端到端驾驶的规划精度。