Learning to Perceive: A Developmental Robotics Approach to Vision and Object Interaction.

摘要

A robot is a true blank slate, awash in sensory information inextricable from its own actions. As such, it can be a powerful tool for investigating the space of problems that an infant, or whatever innate machinery was granted to the infant by evolution, must solve. The key observation is that the environment in which an infant develops is the same as that in which a robot might exist. A robot may have a different view on that environment, through different types or qualities of sensors. A robot may have different capabilities for interacting with the environment, for example by having wheels instead of legs. But, to act autonomously and coherently in the world, like an infant learns to do, a robot must somehow make sense of its sensory information. The goal of this thesis, broadly, is to enable the pneumatic humanoid robot, Diego, to actively perceive the world. Specifically, three problems are addressed. How can a robot: (1) identify interesting information in its sensory input, (2) direct its sensors to best acquire meaningful information, (3) learn to generalize its experience to interact with novel objects? To help answer these questions, this thesis presents: (1) a model of salience that is suitable for active cameras, (2) a model of eye movements based on optimal control, and (3) a framework and robotic implementation for visual perception of the inertial properties of objects.