2025上半年已经过去了，上半年自动驾驶发展的怎么样？业内落地量产在做什么？学术界在关心什么技 术方向？转行具身还是深耕智驾？全都是问题！ 为此，自动驾驶之心采访了国内外数十位学术界和工业界的诸位大佬，和他们一起聊聊对这些问题的看 法，希望对大家有帮助~ 1. 当下前沿技术的发展路线是否已经成熟并适合量产？有哪些难点需要攻克？ 2. 如何看待新兴的技术方向：VLA/VLM、扩散模型、闭环仿真、强化学习、端到端自动驾驶、世界模 型等等 3. 未来自动驾驶技术还有哪些值得探索的新方向？ 4. 深耕自动驾驶 or 投身具身智能，怎么选？ 点击下方 卡片 ，关注" 自动驾驶之心 "公众号 戳我-> 领取 自动驾驶近15个 方向 学习 路线 SJTU 张志鹏 BEV，Occ， E2E， VLA， 这几个词是从2022年到2025年自动驾驶中热词的一个变迁。 BEV和Occ的引入某种程度上都是因为特斯拉推崇纯视觉方案，然后借鉴Lidar和机器人的一些建模方式 去表征整个驾驶环境。近期又出现了一些用3DGS建模3D环境的方式。但是整体来看，自动驾驶中感知 的场景表征的方式还是比较成熟了，各家厂商也验证了这种模式在实际中是走 ...

Core Viewpoint - The article discusses the differences between VLA (Vision-Language-Action) and end-to-end models in the context of autonomous driving, emphasizing the importance of large models and their applications in the industry [2]. Group 1: Job Descriptions and Requirements - Positions related to large model development, including VLA and end-to-end roles, are highlighted, with a focus on skills in fine-tuning, lightweight models, and deployment [2]. - The job of an end-to-end/VLA engineer involves developing and implementing driving systems, optimizing model structures, and constructing high-quality training datasets [6]. - The VLA/VLM algorithm position requires a master's degree in computer science or AI, with 3-5 years of experience in autonomous driving or AI algorithms, and proficiency in VLA/VLM architectures [8][10]. Group 2: Technical Skills and Experience - Candidates are expected to have experience with multimodal large language models, fine-tuning existing models for specific business scenarios, and familiarity with Transformer and multimodal technologies [5]. - Experience in computer vision, trajectory prediction, and decision planning is essential, along with a strong foundation in mainstream technologies and frameworks like PyTorch [9]. - The article emphasizes the need for candidates to have published papers in top conferences or achieved notable results in international competitions [9][11].