绿帽社

Title: Scalable Autonomy in the Era of Foundation Models鈥�

Abstract: Autonomous agents have long been studied in specific applications like robotics and autonomous vehicles, where progress often navigated trade-offs between performance, data efficiency, and generalization. The recent emergence of foundation models, however, is dissolving these traditional barriers, enabling the development of the first truly generalist agents. In this talk, we will discuss the core methods for building the generalist agents, particularly focusing on learning from vast, yet imperfect real-world data. We explore the progression of training paradigms, starting from large-scale imitation learning and advancing to reinforcement learning aligned with a variety of human data, to scaling reinforcement learning to large capacity transformer architectures with suboptimal data. Next, we examine inference-time learning, where agents can adapt to new tasks and environments on-the-fly, without retraining. This adaption capacity, along with self-correction methods are a key step toward self-improving and truly autonomous systems. We will contextualize the societal impact of these advancements using a "Levels of AGI" framework, drawing an analogy to the levels of autonomy in self-driving cars. Finally, we will outline the critical open research problems for the next generation of agentic AI. This includes scalable solutions for online planning and continual learning, the development of robust world models, and ensuring value alignment in open-ended sequential decision-making. We will discuss how new frontiers in bioengineering, computational social sciences, and robotics are not only powerful applications with major societal impact, but also essential drivers for the next generation of foundational agentic systems.

听