Talk 1
Heye Huang
Korea Advanced Institute of Science & Technology (KAIST)
Safe embodied AI and autonomous driving
Talk details
Autonomous systems operating in real-world open environments face a persistent long-tail safety challenge: rare but high-risk scenarios are difficult to anticipate, underrepresented in standard datasets, and often require reasoning beyond pattern recognition. This talk rethinks safe autonomy from the perspective of rare risks and reliable decisions. It presents a unified framework that connects long-tail high-risk scenario data, prior risk knowledge, multimodal semantic understanding, and large language model reasoning to support risk-aware autonomous decision-making in complex traffic environments. By constructing knowledge-augmented risk representations and integrating them with LLM-based reasoning, the framework enables autonomous systems to assess risk sources, understand interaction context, and generate interpretable decisions in rare, extreme, and interaction-intensive scenarios. Ultimately, this work aims to provide new insights into how high-risk data and generative AI can jointly support trustworthy autonomous driving in complex open environments.
