Learning, Memorization and Cognition in an Autonomous Robot Control Architecture

点击翻译按钮获取中文摘要

The ultimate goal of Artificial Intelligence is to work in the wild, messy world, leading to address key challenges such as architecture, learning and evaluation. Mobile robots are evolving from being rigid, two-dimensional navigation platforms, to compliant and interactive machines that can accomplish more complex and sophisticated tasks, assuming that they can exploit their advanced capabilities intelligently. ******To do so, the long term objective of my Discovery research program is the study of how to integrate the required capabilities and decision-making processes so that robots can operate in the real world over long periods of time. ******Consequently, I developed IRL-1, a compliant omnidirectional humanoid mobile robot capable of natural reciprocal interaction (pose, navigation, auditory, visual, language, touch, gesture, affect) with humans. This platform provides an exceptional test bed to address the integration issues that a high-level robot control architecture has to face. The focus of my proposal is to study how data from the robot's external and internal perspectives can be memorized, categorized and learned, to improve the understanding of its interaction dynamic with its operating environment for intelligent and autonomous decision-making. I plan to integrate capabilities that will allow IRL-1 to know where, with whom or what happened when interacting in the world with people, to have it determine what to do next using memory models that can adapt to the situations experienced in the environment and its limited capabilities. A memory management approach will determine which information to keep in the robot's working memory so that online processing constraints are satisfied. An episodic-like memory model will learn to categorize the experiences of the robot, to predict and to influence its intentions for more efficient performance and interaction. Both spatial and episodic-like memory models will complement each other to validate and infer high-level concepts. Taking into consideration the history of events, behavior exploitation and intentions over time will also provide IRL-1 with the ability to self-characterize its experiences based on perceived events and the use of its own control and reasoning processes. ******IRL-1 will be programmed to accomplish tasks with the intent to acquire, iteratively, spatio-temporal knowledge about space and events, and gain knowledge over time from past experiences. Performance will be measured in terms of the added capabilities brought by the high-level control mechanisms, and from the robot's overall ability to operate autonomously in diverse conditions. Experimenting and integrating learning capabilities from multiple modalities will provide essential insights at the foundation of what is required for having service or assistive robots operate autonomously in natural settings, and improve our understanding of our own intelligence.**