About the Talk: Embodied AI as an emerging research topic has been studied in various visuomotor tasks such as indoor navigation and autonomous driving. Most embodied AI studies are conducted in fixed simulation environments, where the generalizability and safety of the autonomous agents in unseen complex scenes remain questionable. In this talk, I will give an overview of the research works in my lab for building generalizable AI for machine autonomy, from the aspects of training data/environment, visual representations, and learning pipelines. First, I will introduce our effort in building the MetaDrive driving simulator by incorporating the capability of generating infinite novel traffic scenarios. Then I will talk about learning generalizable neural representations for decision-making from hours of uncurated YouTube driving videos. Finally, I will discuss our work on Human-in-the-loop reinforcement learning for safe exploration and inference.
- MetaDrive simulator: https://metadriverse.github.io/metadr...
- ECCV’22 paper: Learning to drive by watching YouTube videos: Action-Conditioned Contrastive Policy Pretraining. https://arxiv.org/pdf/2204.02393.pdf
- ICLR’22 paper: Efficient Learning of Safe Driving Policy via Human-AI Copilot Optimization. https://decisionforce.github.io/HACO/

About the Speaker: Bolei Zhou is an Assistant Professor in the Computer Science Department at the University of California, Los Angeles (UCLA). He earned his Ph.D. from MIT in 2018. His research interest lies at the intersection of computer vision and machine autonomy, focusing on enabling interpretable and trustworthy human-AI interaction. He has developed a number of widely used interpretation methods such as CAM and Network Dissection, as well as computer vision benchmarks Places and ADE20K. He is an associate editor for Pattern Recognition and has been area chair for CVPR, ECCV, ICCV, and AAAI. He received MIT Tech Review's Innovators under 35 in Asia-Pacific Award. More about his research is at https://boleizhou.github.io/ .