In order to solve complex tasks, intelligent agents learn to perform a variety of skills. These skills abstract key information necessary for agents to act in real-world environments with high-dimensional state and action spaces. The goal of the Smart Practice project is to develop learning algorithms that can efficiently discover, compose, and execute a curriculum of skills from demonstration data.
- Yuqing Du, UC Berkeley, https://yuqingd.github.io/
- Pieter Abbeel, UC Berkeley, https://people.eecs.berkeley.edu/~pabbeel/
- Aditya Grover, Facebook AI Research, http://aditya-grover.github.io/
Learning autonomous agents that can perform complex, high-dimensional tasks is a long standing goal of artificial intelligence. In this project, our goal is to learn primitives necessary for such agents from demonstration data. Demonstration data from experts e.g., humans can directly encode intricate behaviors for complex tasks such as self-driving.
Existing work in this space mostly relates to learning imitation policies from demonstrations of near-optimal expert behavior. The supervision here is in the form of low-level actions provided for each state. The primary goal of this project is to build on recent advances in weakly-supervised machine learning to design and analyze learning algorithms that directly learn higher-level spatio-temporal abstractions, or skills. The kinds of supervision we will explore could be in many different forms: ranking of demonstrations, human preferences, additional modalities, observation-only data streams (e.g., videos).
Finally, post-learning of latent skills, we are also interested in learning a curriculum for the order of skills that enables fast and accurate learning across a range of tasks.