Learning Macro Actions from Instructional Videos Through Integration of Multiple Modalities

摘要

We propose an architecture for a system that will "watch and listen to" an instructional video of a human performing a task and translate the audio and video information into a task for a robot to perform. This enables the use of readily available instructional videos from the Internet to train robots to perform tasks instead of programming them. We implemented an operational prototype based on the architecture and showed it could "watch and listen to" two instructional videos on how to clean golf clubs and translate the audio and video information from the instructional video into tasks for a robot to perform. The key contributions of this architecture are: integration of multiple modalities using trees and pruning with filters; task decomposition into macro-tasks composed of parameterized task-primitives and other macro-tasks, where the task-primitive parameters are an action (e. g., dip, clean, dry) taken on an object (e. g., golf club) using a tool (e. g., pail of water, brush, towel); and context, for determining missing and implied task-primitive parameter values, as a set of canonical task-primitive parameter values with a confidence score based on the number of times the parameter value was detected in the video and audio information and how long ago it was detected.