PhD student, UC Berkeley
Learning to Learn: Enabling Robot Versatility with Adaptable Deep Representations
In order for robots to be capable of performing a wide range of jobs in unstructured settings, they must be able to acquire new skills quickly and efficiently while adapting to dynamic environments. High-capacity models such as deep neural networks can enable a robot to represent complex skills, but learning each skill from scratch requires large amounts of experience. We present a meta-learning method that enables a robot to learn much more efficiently, allowing it to learn to interact with new objects and reach new goals from tens of trials or just a single demonstration. Our experiments on both simulated and real robot platforms demonstrate the ability to scale to complex skills, including those requiring vision, with minimal data. Our approach additionally achieves state-of-the-art performance on few-shot image classification and can be applied to learning concepts from only positive examples.
Chelsea Finn is a PhD student in Computer Science at UC Berkeley, studying machine learning for perception and control of embodied systems. She is interested in how learning algorithms can enable robots to acquire common sense, allowing them to learn a variety of complex sensorimotor skills in real-world settings. During her PhD, she has developed deep learning algorithms for concurrently learning visual perception and control in robotic manipulation skills, inverse reinforcement methods for scalable acquisition of nonlinear reward functions, and meta-learning algorithms that can enable fast, few-shot adaptation in both visual perception and deep reinforcement learning. Chelsea received her Bachelors degree in Electrical Engineering and Computer Science at MIT. She has also spent time as an intern at Google Brain, working on self-supervised robot learning algorithms using deep predictive models with data from several robot arms. Her graduate research has been supported by an NSF graduate fellowship. With Sergey Levine and John Schulman, she also designed and taught a course on deep reinforcement learning, with more than a thousand followers online.
For links to papers, videos, and open-sourced code and data, see: https://people.eecs.berkeley.edu/~cbfinn/