Wave 公司 CEO Alex Kendall 讨论自动驾驶从经典机器人学向端到端深度学习的范式转变，强调泛化能力和安全性。
Alex Kendall, CEO of Wave, discusses the paradigm shift from classical robotics to end-to-end deep learning in autonomous driving, emphasizing generalization and safety.
要点 · TL;DR
端到端神经网络在自动驾驶中优于手工编码的堆栈。 End-to-end neural nets outperform hand-coded stacks in autonomous driving.
世界模型实现推理和安全行为，如在盲弯处试探前行。 World models enable reasoning and safe behaviors like nudging on blind turns.
语言集成有助于表示、人机交互和内省。 Language integration aids representation, human interaction, and introspection.
核心观点 · Key points
端到端深度学习在自动驾驶上优于手工编码的堆栈。 End-to-end deep learning outperforms hand-coded stacks for autonomous driving.
泛化是关键：AI 必须通过多样化数据和世界模型处理未见场景。 Generalization is key: AI must handle unseen scenarios via diverse data and world models.
世界模型实现推理和涌现的安全行为，如在盲弯前试探前进。 World models enable reasoning and emergent safe behaviors like nudging forward on blind turns.
与 OEM 合作可在不改装硬件的情况下实现规模化部署。 Partnering with OEMs allows scalable deployment without retrofitting hardware.
语言集成改善表征，支持人机交互，并帮助内省。 Language integration improves representation, enables human interaction, and aids introspection.
通过课程学习和合成数据实现数据效率对成本和速度至关重要。 Data efficiency through curriculum learning and synthetic data is critical for cost and speed.
反共识 · Contrarian takes
端到端神经网络如今借助现代工具是可解释的，与过去的看法相反。 End-to-end neural nets are interpretable today with modern tools, contrary to past beliefs.
基于规则和学习的混合系统往往兼得两者之短。 Hybrid rule-based and learned systems often get the worst of both worlds.
纯摄像头可达人类水平，但超人类安全需要雷达和激光雷达。 Camera-only can reach human level, but beyond-human safety needs radar and lidar.
自动驾驶通过扩展当前方法即可解决，无需新突破。 Autonomous driving is solved by scaling current approaches, not requiring new breakthroughs.
AV 3.0 可能通过网状网络将智能移至车外，消除红绿灯。 AV 3.0 may move intelligence outside the car via mesh networks, eliminating traffic lights.
完美的模拟器等于解决了自动驾驶，类似于 AlphaGo 的方法。 Perfect simulator equals solved self-driving, similar to AlphaGo's approach.
本期章节 · Chapters（共 23）
引言与 AV 1.0 vs 2.0Introduction and AV 1.0 vs 2.0
Wave 方法：传感器输入到运动输出Wave's approach: sensor input to motion output
逆向思维起点与可解释性Contrarian beginnings and interpretability
前后对比：AV 1.0 vs 2.0 堆栈Before and after: AV 1.0 vs 2.0 stacks
传感器架构与分布Sensor Architecture and Distribution
泛化与全球部署Generalization and Global Rollout
推理与世界模型Reasoning and World Models
数据多样性与来源Data Diversity and Sources
数据效率与世界模型Data Efficiency and World Models
公司文化与汽车行业Company Culture and Automotive Industry
向 OEM 学习Learning from Auto OEMs
与 OEM 的市场路径Path to Market with Auto OEMs
传感器融合辩论Sensor Fusion Debate
跨车型模型适配Model Adaptation Across Vehicles
边缘案例与全球适配Corner Cases and Global Adaptation
语言模型集成Integration of Language Models
内省与车载计算Introspection and Onboard Compute
从自动驾驶到具身 AIFrom Autonomous Driving to Embodied AI
物理 AGI 的研究突破Research Breakthroughs for Physical AGI
AV 3.0 愿景AV3.0 Vision
AV 3.0 未来：超越汽车的智能Future of AV 3.0: Intelligence Beyond the Car