Robust Learning of Visual Behaviors

点击翻译按钮获取中文摘要

The Robust Learning of Visual Behaviors for robots will bring more effective helpers for Canadian society and industry. Rather than requiring a large team of engineers and programmers, this research proposes a general-purpose set of learning tools that allow a robot to improve its own behavior over time, by trial-and-error. This will not only make our robot helpers easier to use, it will allow them to adapt their behaviors appropriately, as the needs of human users and challenges of the environment change. There are three aspects to the work in this proposal. First, we will work on the core Reinforcement Learning technology, which produces software that acts as the robot's "brain" in this context. There, our technologies are continually improving, and only lately, it has been possible to use Reinforcement Learning to train software agents to beat humans at boardgames like Chess and Go as well as many video games. The same technology is now being extended to allow robots to learn in this way. Second, we will connect the robot's visual sensing to this learning system, so that robots guided by camera can learn tasks that involve eye-hand coordination, among other examples. Finally, we seek to make it easy for users to instruct robots to learn in a way that is most useful to human users. Interactive Behavior Authoring will provide ways for the robot and human to enter a dialogue where the human gives information on what is intended, and the robot queries for more information where it can benefit most. The core technology in this proposal, Reinforcement Learning, is already broadly used in Canadian industries. Technology and AI companies use this to train dialogue agents that answer phones and improve search engines. Medical researchers have used RL to investigate new courses of treatment. The most relevant uses to this proposal involve the automation of moving physical bodies. Robot arms are often present in today's factories and warehouses, assisting human workers with the most dull and repetitive tasks. The new work in this proposal will extend the set of places and tasks where robots can be deployed with Reinforcement Learning. We will target driving in forests so that timbre harvesting and transport can benefit. Farms, mines and watersheds are also areas where we could make great use of robotics, and this proposal will produce useful technologies for increasing efficiency in each of these domains. Finally, as this proposal is written during the time of the COVID pandemic, there are many humans sheltering at home and isolated. There is great potential for Reinforcement Learning to produce virtual and physical companions who are immune from the virus, and therefore of great prospect as housemates until the new normal returns.