Even before the deep learning revolution, the landscape of artificial intelligence (AI) was already changing drastically in the 90s. Embodied intelligence, it was proposed, must play a crucial role in the design of intelligent machines. This new wave was inspired by what is today known as Embodied and Enactive Cognitive Science or E-Cognition, which considers that cognitive activity does not reduce to the intellectual capacities of agents being able to represent their environments. E-cognition set AI and robotics in a new direction, in which intelligent machines are required to interact with the environment, and where this interaction does not reduce to explicit representations or prespecified algorithms.
These ideas revolutionized the way we think about intelligent machines and cognition, but these theoretical advances are only partially reflected in modern approaches to AI and machine learning (ML). Despite deeply impressive achievements, AI/ML still struggles to recapitulate the kinds of intelligence we find in natural systems, whether we are considering individual insects (e.g. simultaneous localization and mapping), or swarm behaviour (e.g. forum sensing and ensemble inferences), and especially the kinds of flexibility and high-level reasoning characteristic of human cognition.
How can we build embodied intelligent systems capable of adapting to dynamically changing environments with coherent purposes and sensibly prioritized goals, choosing which objectives to pursue based on environmental conditions and particular contexts (cf. frame problem)? That is, how can we ensure that the complexity of intelligent systems mirror the complexity of their environments, potentially without centralized control structures or explicit representations?
Meta-learning may be particularly notable in providing a bridge between more enactivist and traditionally cognitivist perspectives on intelligence. However, while aspects of meta-learning appear to spontaneously emerge from recurrent systems exposed to iterative tasks, useful knowledge-transfer requires substantially overlapping task-structure across epochs. To what extent can richly-structured, non-trivially controllable embodiments provide these overlapping task-demands required for sustainable development of increasingly sophisticated capacities with continual learning?
We invite contributions on the nature of intelligence and its potential recapitulation by artificial systems. Relevant questions include (but are not limited to) the following:
• How can we build embodied intelligent systems that work in a dynamically changing environment?
• What counts as embodiment for AI?
• What kind of embodiments and environmental embeddings are required for what kinds of cognitive capacities (e.g. robust inference, lifelong learning, flexible adaptation, etc.)?
• What are the limits of simulated embodiments and environments, how can these limits be overcome?
• Are there aspects of embodied cognition that can be achieved by other means (e.g. deployment of large-scale computing systems), or are there aspects of intelligence that fundamentally require an enactivist approach?
• In developing advanced AI, what potentialities and limitations might be suggested by an embodied perspective?
• What is required for flexibly adaptive robots that we can deploy in real-world situations, including those involving human beings?
Even before the deep learning revolution, the landscape of artificial intelligence (AI) was already changing drastically in the 90s. Embodied intelligence, it was proposed, must play a crucial role in the design of intelligent machines. This new wave was inspired by what is today known as Embodied and Enactive Cognitive Science or E-Cognition, which considers that cognitive activity does not reduce to the intellectual capacities of agents being able to represent their environments. E-cognition set AI and robotics in a new direction, in which intelligent machines are required to interact with the environment, and where this interaction does not reduce to explicit representations or prespecified algorithms.
These ideas revolutionized the way we think about intelligent machines and cognition, but these theoretical advances are only partially reflected in modern approaches to AI and machine learning (ML). Despite deeply impressive achievements, AI/ML still struggles to recapitulate the kinds of intelligence we find in natural systems, whether we are considering individual insects (e.g. simultaneous localization and mapping), or swarm behaviour (e.g. forum sensing and ensemble inferences), and especially the kinds of flexibility and high-level reasoning characteristic of human cognition.
How can we build embodied intelligent systems capable of adapting to dynamically changing environments with coherent purposes and sensibly prioritized goals, choosing which objectives to pursue based on environmental conditions and particular contexts (cf. frame problem)? That is, how can we ensure that the complexity of intelligent systems mirror the complexity of their environments, potentially without centralized control structures or explicit representations?
Meta-learning may be particularly notable in providing a bridge between more enactivist and traditionally cognitivist perspectives on intelligence. However, while aspects of meta-learning appear to spontaneously emerge from recurrent systems exposed to iterative tasks, useful knowledge-transfer requires substantially overlapping task-structure across epochs. To what extent can richly-structured, non-trivially controllable embodiments provide these overlapping task-demands required for sustainable development of increasingly sophisticated capacities with continual learning?
We invite contributions on the nature of intelligence and its potential recapitulation by artificial systems. Relevant questions include (but are not limited to) the following:
• How can we build embodied intelligent systems that work in a dynamically changing environment?
• What counts as embodiment for AI?
• What kind of embodiments and environmental embeddings are required for what kinds of cognitive capacities (e.g. robust inference, lifelong learning, flexible adaptation, etc.)?
• What are the limits of simulated embodiments and environments, how can these limits be overcome?
• Are there aspects of embodied cognition that can be achieved by other means (e.g. deployment of large-scale computing systems), or are there aspects of intelligence that fundamentally require an enactivist approach?
• In developing advanced AI, what potentialities and limitations might be suggested by an embodied perspective?
• What is required for flexibly adaptive robots that we can deploy in real-world situations, including those involving human beings?