Coursera Spring Sale
40% Off Coursera Plus Annual!
Grab it
Explore the fundamental parallels between robotics and AI agents in this 18-minute conference talk that draws critical lessons from the self-driving car industry. Discover how the robotics field initially misunderstood the relative difficulty of perception versus planning, spending 8-10 years learning that planning was actually the harder challenge, and examine how this same pattern is emerging in today's agent development. Learn why predictive models differ fundamentally from action models and understand why perfect reasoning doesn't guarantee effective execution in real-world scenarios. Delve into key robotics concepts including DAgger (Dataset Aggregation), Markov Decision Processes (MDPs), simulation techniques, and offline reinforcement learning, and see how these principles directly apply to building scalable agent systems. Gain insights into why infrastructure capabilities often matter more than model performance alone when deploying agents in production environments. The presentation draws from extensive experience across computer vision, natural language processing, deep learning recommender systems at YouTube, transformer-based planning models for autonomous vehicles at Waymo, and current work applying large-scale reinforcement learning and simulation techniques to coding agents.
