[T@W Intro] Corina Gurau - World Models for Autonomous Driving скачать в хорошем качестве

[T@W Intro] Corina Gurau - World Models for Autonomous Driving

Corina Gurau (Air Street) is talking at Zeta Alpha's Transformers at Work 2023 and her talk will be focused on World Models for Autonomous Driving, research that she worked on in while at Wayve. Models that drive autonomous vehicles can benefit from world models. These help them learn more general representations about how the world works, and they should speed up the process of learning to plan, and generalize better in different environments. The architecture for learning a world model is based on a model called Dreamer, that's been pretty successful for planning. Humans learn a lot of knowledge about the world just by passively passively observing it, and when we learn end-to-end models for driving, a lot of the representational power of the model goes into perception, and central in this work is a world model that focuses on the visual understanding of the world. If we learn really good representations, we could we could leverage things like maybe doing reinforcement learning inside of the world model, and that should that should help with learning driving policies. At the moment, learning rewards or this kind of value assigned to a world model state is really tricky for autonomous driving or in the real world. It's important that the the role model that we use has spatial understanding, and this should also make for more interpretable world models, which is one of the criticisms of end-to-end learning To hear the full talk, sign up for Zeta Alpha's Transformers at Work 2023: https://www.zeta-alpha.com/events/tra...