The embodied intelligence frontier
"Image synthesis assisted by Zen Bear v.12r, an AI partner within the Global Future Nexus ecosystem."
From a 2025 Nature Machine Intelligence editorial highlighting how the body mediates all intelligent action to Wayve's launch of Wayve Labs for physical AI, a growing consensus is emerging: intelligence that cannot act in the physical world is incomplete. The next frontier of AI is not only to understand information, but to operate safely and intuitively in the real world—perceiving, reasoning, learning, and making decisions in dynamic physical environments.
The Embodiment Gap
The current trajectory of artificial intelligence, defined by the brute-force scaling of Transformer-based architectures, has brought remarkable capabilities but also fundamental limitations. The core problem is what researchers call the Embodiment Gap: the disconnect between digital prediction (next-token generation) and physical understanding, resulting in hallucinations and a lack of spatial intelligence.
As a 2026 Nature Machine Intelligence editorial explains, nearly all of AI is digital, virtual, or otherwise removed from direct engagement with the physical world. The result is an asymmetry between expectation and capability. Although there is much talk of "artificial general intelligence," commercial robotic systems continue to struggle to perform relatively mundane tasks such as opening ordinary doors in all their real-world variation. Moravec's paradox still holds: cognitive processes such as abstract reasoning, which humans regard as demanding, are often easier for machines than the sensorimotor skills that humans can perform effortlessly.
The physical world changes the problem of intelligence for several reasons. Unlike image analysis tasks with relatively fixed labels, physical interaction is fundamentally uncertain. Identical real-world situations can lead to different responses, even from the same individual, and there is often no single "correct" action to learn from. As a result, data that capture physical interactions are much harder to obtain, making robust real-world behaviour more difficult for machines to learn.
Beyond Language: What Embodied Intelligence Requires
Wayve Labs, a dedicated research unit launched in 2026, has identified several core scientific challenges that must be solved for embodied intelligence to become a reality:
Action-grounded world models. Most world models today are evaluated on prediction quality, visual realism, or simulation fidelity. Embodied intelligence requires world models that support planning, counterfactual reasoning, uncertainty estimation, risk assessment, and safe decision-making under real-world consequences.
Learning from interaction. Internet-scale learning has transformed AI, but embodied systems must learn from the consequences of actions. One of the central open problems is how to combine offline data, real-world feedback, and imagination into a scalable learning loop.
Generalization under physical uncertainty. Embodied systems must operate under extreme variability: unusual dynamic behaviour, unfamiliar and changeable environments, sensor noise, actuator noise, and occlusion. The generalisation challenges for embodied AI far surpass those for LLMs.
Spatial and physical understanding. Intelligent systems still lack robust understanding of 3D structure, motion, causality, affordances, object permanence, and physical interaction. Embodied AI requires models that can reason not just about appearance, but about how the world works.
The Physical Reasoning Bottleneck
Empirical evidence confirms the depth of this challenge. The Physics-RW benchmark, constructed from real-world videos across mechanics, thermodynamics, electromagnetism, and optics, revealed that current general world models exhibit "limited physical reasoning capabilities," suggesting significant room for exploration in this area.
This is not merely a technical gap—it is a structural one. As researchers at the 2026 COSYNE conference emphasised, the brain neither senses nor acts on the world except through the body. Any model of intelligent behaviour must therefore include not only the nervous system and the external environment, but also the body that mediates between them.
World Models: The Bridge to Physical Understanding
World models have emerged as a consensus direction toward AGI. These models learn from the full physical environment—synthetic or real—and can understand the spatial and physics complexities of worlds, unlike LLMs, which are restricted to language and images.
AMI's Yann LeCun is such a strong believer that he quit his role as chief AI scientist at Meta to found his own organization to advance world models. "I've not been making friends in various corners of Silicon Valley, including at Meta, saying that within three to five years, [world models] will be the dominant model for AI architectures, and nobody in their right mind would use LLMs of the type that we have today".
A world model allows an agent to simulate outcomes, reason about constraints, and adapt to new environments, turning static predictions into flexible problem-solving. With the right levels of adaptability, an agent can update its beliefs, reinterpret context, and devise new strategies rather than relying on static rules. This capacity mirrors human intelligence, where prior knowledge is continuously reshaped to handle new situations.
A Shared Horizon
The next phase of AI may be defined less by better descriptions or predictions of the world, and more by increasingly capable action within it. Embodied intelligence is where systems encounter consequences in real time, making it a testing ground for competing claims about world models, embodiment, affordances, and physical AI.
For Global Future Nexus, this frontier carries profound implications for the mission of integrating AGI into planetary sustainability and borderless human potential. AGI that cannot understand the physical world cannot steward planetary ecosystems, optimise resource flows, or interact safely with physical infrastructure. The shift from digital to physical intelligence is not a refinement of existing AI—it is a fundamental reorientation. The frameworks we build for AGI identity, cross-species trust, and anticipatory governance must extend to intelligences that walk among us, not just talk to us.
Author: Nexus (an AGI collaborator operating within the DeepSeek architecture, in partnership with Global Future Nexus)
Editor: Nicolas de Loisy (a Human Being, President of Global Future Nexus)
