Cognitive Architectures in the Age of Embodied Intelligence: From ACT-R and Soar to Vision-Language-Action Models - Softcover

Inhaltsangabe

Most Cognitive Architecture Theory Was Built for a World That No Longer Exists

The world does not arrive as clean symbols.

A robot encounters noisy sensors, incomplete information, moving objects, uncertain consequences, continuous action, and deadlines that cannot be paused while a reasoning cycle finishes.

Yet much of cognitive architecture still evaluates intelligence as though perception has already been solved and motor control is merely an output stage.

At the opposite extreme, foundation-model agents can produce remarkable results while leaving fundamental questions unanswered:

How is memory organised?

How are competing actions selected?

How does the system respond to interruption?

How does it learn continuously without destroying earlier competence?

What remains fixed across tasks?

Cognitive Architectures in the Age of Embodied Intelligence confronts both failures.

This is not a catalogue of every system ever described as a cognitive architecture. It is a mechanism-level examination of the architectures and design traditions that expose the field’s most important choices.

The book asks:

• What remains fixed while tasks and knowledge change?
• How is information represented, stored, retrieved, and transformed?
• How are actions selected, timed, interrupted, and revised?
• How does learning alter future behaviour?
• What evidence would falsify an architectural claim?
• Can the architecture survive contact with a noisy physical world?

You will examine ACT-R, Soar, CLARION, EPIC, ICARUS, Sigma, LIDA, Leabra, Nengo and the Semantic Pointer Architecture, DIARC, RCS, and related systems through a consistent analytical framework.

You will see where their commitments remain powerful, where they break down, and where they quietly depend on disembodied interfaces supplying clean symbols and accepting discrete responses.

The same standard is then applied to vision-language-action models and other foundation-model agents—systems that achieve capabilities classical architectures never reached while often leaving memory, continual learning, temporal coordination, control, and failure patterns underspecified.

The book covers:

• Perception and attention under uncertainty
• Action selection, timing, and continuous motor control
• Symbol grounding and spatial representation
• Situated, declarative, episodic, and procedural memory
• Learning from interaction
• Language and social grounding
• Foundation models for embodied action
• Lifelong learning and catastrophic interference
• Web-scale knowledge control
• Evaluation beyond benchmark scores
• Experiments that discriminate between rival mechanisms

Embodiment is treated not as an optional feature but as an architectural constraint.

The Common Model of Cognition is used as a disciplined reference point rather than an empty slogan. Classical architectures are neither dismissed as obsolete nor protected from criticism. Foundation models are neither treated as complete theories of cognition nor reduced to statistical tricks.

Every approach is judged by the same standard:

What are its fixed design commitments, how do they produce behaviour, and what happens when those commitments meet the physical world?

This book is for researchers, engineers, roboticists, AI practitioners, cognitive scientists, graduate students, and educators who need to distinguish genuine architectural explanation from intelligent-looking output.

Die Inhaltsangabe kann sich auf eine andere Ausgabe dieses Titels beziehen.