Artificial intelligence, cognition and learning

AI and learning are critical to robot cognition; that is, the way robots acquire knowledge and understanding of the world around them, then reason and make decisions. Cognition is the main challenge preventing robots from operating autonomously in unstructured environments. Robots are not yet as flexible or adaptable as humans and currently lack human-like intuition and reasoning abilities [AAS18]. Meta-learning, or learning how to learn new things, is a critical new AI capability that must be applied to robots [SM18]. To be able to learn on the fly, adapting to dynamic and uncertain environments and to understand their own limitations is critical for robots to find wider applications than those currently observed.

Robotic systems that know how to interact (with people, other robots and their environment), how to seek help, how to recover from failure, and how to become smarter are needed. Eventually, AI and learning needs to give robots the ability to model their own components and operations so that robots can adapt and evolve [SM18]. To achieve this, AI needs to be able to ‘self-learn’ complex tasks with a minimum of initial training data. Current machine-learning systems are data-intensive, requiring massive amounts of data to learn tasks, yet are unable to readily apply that knowledge to other domains. The complex and changing nature of the real world makes it difficult to build robust systems that readily learn, but the application of neuroscience to the development of AI systems is a start [SM18]. Once robots have successfully mastered a task they then rely on common task libraries to be able to share their learnings with other robots, building a giant store of robotic knowledge that is only starting to be appreciated. Sharing data is easy but a key challenge is how this learning is represented.

One of the greatest challenges in robot cognition is to develop cognitive capabilities that can push back the level of structure required in the operating environment, until robots can operate safely and effectively in unstructured environments. Breakthrough research directions include finding universally applicable approaches to transfer learning, being able to operate in both symbolic and non-symbolic logics, developing compact representations of the world that admit various forms of reasoning, and developing cognitive architectures that can reliably make sense of the surrounding environment and then take appropriate action [AAS18].