Our long-term vision for advanced autonomy is based on effecting a step change in terms of how we develop, integrate, adapt, and interact with AI systems – where autonomy is understood as the ability of software or hardware systems to perform complex tasks delegated to them competently, safely, and in alignment with human expectations.
We expect these autonomous agents (e.g. robots, softbots, conversational assistants etc), to be able make complex decisions and act over and above extracting knowledge from data and reasoning about it.
We expect that achieving these levels of autonomy will require a concerted effort to develop new methods by working across all areas of AI, where key new advances will likely come from a range of methods and subfields that many researchers are already working on, including:
New ways of dealing with data
Moving beyond the currently practice of training systems on large amounts of supervised data will require advancing techniques for learning from limited supervision. New approaches for one-/few-shot learning, combining different data sources, new forms of data, and new ways of using high-level knowledge in combination with observational data are needed.
New ways of assembling systems
To overcome the focus of current systems on very narrow, often single-shot tasks, and enable on-the-fly combination of individual components in novel modular and expandable systems, methods such as meta-learning, multiagent coordination, or neuro-symbolic architectures will likely be needed.
New ways of adapting systems
Enabling systems to adapt to different contexts of use will require exploring new forms of adaptation, for example through targeted human instruction and demonstration, but also to enable adaptation to the needs of diverse user populations, using novel lifelong learning, task generalisation/ abstraction, and crowdsourcing.
New ways of interacting with systems
Increasingly complex AI systems will likely be programmed and configured by developers and users at runtime in novel, more natural, and incremental ways. Adjustable autonomy, mixed-initiative techniques, human-AI interaction/communication, interpretability, and new programming techniques will play an important role here.
New ways of understanding and using autonomy
Whatever levels of autonomy we strive for, future AI will have to complement human activity meaningfully and safely, exploiting the strengths of artificial and human intelligence. Ensuring we pursue responsible and beneficial innovations will require deep exploration of the ethical, human, and societal aspects of autonomy.