The new perspective article, published in Frontiers in Neurorobotics by Dario Pasquali, Francesco Rea, Pietro Morasso, Giulio Sandini and Alessandra Sciutti, takes aim at the dominant paradigm in robotics and artificial intelligence. Foundation models, vision-language-action systems, memory-augmented agents and world models have produced impressive advances in perception, reasoning and embodied control. Yet the authors point out that nearly all of these systems rest on the same assumption: that ever-larger datasets and models will eventually yield open-ended adaptation. That assumption carries costs. Scaling data collection and model size is already raising environmental, economic and societal concerns, and the resulting knowledge offers little guarantee of generalizing to the specific nuances of a particular home, clinic or workplace.

There is also a subtler problem, one the authors call an epistemological fault line between humans and machines. In large language model based robotic systems, knowledge is encoded as static parameters that carry little information about when it was acquired, under what conditions, or how it changed through interaction. Time is treated as a background variable. Generated explanations do not necessarily reveal the processes that produced a decision, and local adaptation does not reorganize the system as a whole. For a non-expert user, this makes a robot’s behavior hard to predict, its motivations hard to trust, and its competences hard to teach in context-sensitive ways. Biological cognitive agents, by contrast, learn continuously, developmentally, and through their bodies.

The researchers propose a shift they frame as Artificial Cognition, a self-developed form of knowledge that supports an agent’s ongoing, survival-oriented adaptation, as distinct from Artificial Intelligence’s emphasis on pre-specified competence. Two foundational conditions anchor their argument. First, experience must be continuous, situated and agent-centered: not a passive stream of data but the ongoing result of embodied interaction, progressively internalized through memory, simulation and prediction. Second, the agent should be what philosophers Aaron Sloman and Jackie Chappell call altricial, like a human infant, designed from the start to develop by relying not only on its own built-in competences but on interaction with other social, cognitive agents. Cognition, in this view, is scaffolded from the outset by an environment populated by intelligent others, echoing Lev Vygotsky’s insight that human higher cognitive functions arise through prolonged embodied and social interaction.

Technically, the team conceives cognitive agents as dynamical systems of parallel loops evolving over time. Behavior is not triggered by external requests but emerges from evolving exchanges among internal states, embodied action and social interaction, all supported by memory. Fast loops capture immediate sensorimotor contingencies, while slower loops integrate experience, stabilize behavioral patterns and support anticipation of delayed outcomes. Rather than adopting the discrete System 1 and System 2 processing of dual-process theory, the authors hypothesize that the foundational ability of cognitive agents is to generate, sustain, close and make sense of multiple parallel loops, which through development expand, get reused and multiply across broader bodily, physical, social and temporal horizons. The same underlying hardware should support both overt action and covert internal simulation, enabling the agent to act, predict and anticipate others using shared embodied mechanisms.

Three pillars hold this framework up: embodiment, memory and values. The body grounds cognition in interaction, producing internal models of the body, the environment and other agents that support simulation and anticipation. Memory gives cognition its temporal depth, but crucially the authors treat it as an active, reconstructive developmental process rather than a passive store. Experiences are progressively encoded, reorganized and reshaped; consolidation, compression, reinforcement and decay transform short-term traces into long-term structures; and retrieval reinterprets past episodes in light of the current context, allowing the same experience to acquire new meaning over time. Values, finally, define what is relevant and worth pursuing. Biological agents start with primitive value structures tied to survival and social stimuli, then develop intrinsic motivations such as curiosity, and eventually stable but revisable preferences, in other words, personality. Values and memory form a self-reinforcing loop: values determine which experiences are attended and remembered, while memory continually reshapes values through reinterpretation. Cognition, the authors write, is not only a matter of knowing but of caring.

ISEED is the first implementation of these ideas, developed within the iCog Initiative, a series of multidisciplinary workshops aimed at designing a cognitive architecture for the iCub humanoid robot. The current embodiment is the ICubHead, a six-degree-of-freedom actuated robot head mounted on a stationary upper body, equipped with RGB cameras and an external stereo microphone. The system has no explicit task. Its behavior is driven by a curiosity-inspired proto-motivation: attention toward environmental deviations that disrupt perceptual consistency and may reveal socially meaningful patterns. The architecture comprises four modules. A Multimodal Perception module processes real-time video and audio, extracting five features to which humans are sensitive from early development: the number of faces, mutual gaze, motion, illumination and stereo audio energy. An Embodied Behavior module alternates phases of stillness and gaze exploration, seeking and tracking human faces in a minimal form of active perception. An Episode Segmentation module synchronizes sensory data and joint states into timestamped events, then groups them into episodes, creating a new episode whenever perceptual features deviate significantly, at least two standard deviations from the mean of each modality, from the statistics of the current one, with changes in the number of faces weighted more heavily than motion or audio as a minimal proto-value system. Finally, an Associative Memory encodes each episode into a 528-dimensional multimodal embedding combining outputs of the pretrained SlowFast video network and VGGish audio network, then performs daily consolidation, reorganizing episodes into associative clusters through unsupervised dimensionality reduction and density-based clustering. These clusters act as emergent social contexts, enabling recognition of familiar versus novel situations and a primitive form of temporal expectation about which context is likely at a given time.

The experimental milestones are striking for their simplicity. Deployed in the institute’s open office, the robot collected 449 episodes over 24 hours across 4 days; at the demo session of an international conference it collected 2,304 episodes over 15 hours across 3 days. From the first days of observation, the architecture autonomously organized its experience into two clusters, learning to distinguish lively episodes, rich in faces, mutual gaze, motion and audio, from calm episodes with little or no human activity, without any supervised labels. The associative structure remained coherent even when experiences from both environments were merged. Recognition accuracy in the office improved steadily with experience, from 0.23 on the first day to 0.87 and then 0.93, and expectation consistency rose from 0.24 to 0.95 and 0.96. At the conference, recognition and temporal stability remained reasonably good, around 0.83 and 0.73 accuracy, but expectation consistency was lower, likely because of the environment’s higher variability and differing recording schedules. Performance, in other words, tracks the regularity and predictability of the experienced environment, exactly as one would expect from a memory grounded in first-person developmental experience rather than pre-labeled abstractions.

The authors are careful to position ISEED not as a complete cognitive architecture but as a minimal testbed for studying the early emergence of Artificial Cognition. The next steps are mapped out in detail. The current implementation processes perception, segmentation, memorization and recognition sequentially; a fuller realization of the dynamical-loops framework requires these processes to unfold continuously and interact over multiple timescales, with memory and motivational dynamics following their own temporal profiles of consolidation, reinforcement and decay. The team is also working to close the cognitive loop, letting the robot decide how to act based on perceived context, past experience and its own values, with context recognition becoming continuous and online rather than available only after an episode ends. To study these developments they are exploring a complementary paradigm they call the Robotic Nursery, a developmentally inspired environment such as a floor playmat or high-chair setup in which an artificial agent can progressively acquire richer sensorimotor, social and contextual experience under controlled but naturalistic conditions.

The broader implications challenge the current rush toward general-purpose robotic intelligence, in which a single pretrained model is expected to provide a universal solution across tasks, environments and users. In homes, clinics, schools and workplaces, the authors argue, effective behavior depends on adapting to specific people, routines, expectations and exceptions through continuous embodied interaction. A developmental architecture lets competence emerge locally from first-hand experience, potentially yielding robots that are more robust to context variability, more sensitive to individual users, more teachable through everyday interaction, and less dependent on centralized data collection. Growing every robot from scratch is technically unfeasible, so the long-term goal is not to abandon pretraining but to treat it as an evolutionary background that bootstraps always-on developmental mechanisms. The real challenge for robotics, the team concludes, is not to build agents that know everything in advance, but agents that can progressively become what their specific social environment requires them to be.

Subject of Research: An experience-driven, always-on cognitive architecture enabling humanoid robots to develop personal knowledge through embodied social interaction

Article Title: ISEED: an always-on cognitive architecture for experience-driven interaction-oriented humanoid robots

Article References: ISEED: an always-on cognitive architecture for experience-driven interaction-oriented humanoid robots. (n.d.). https://doi.org/10.3389/fnbot.2026.1920808

Image Credits: AI Generated

DOI: 10.3389/fnbot.2026.1920808

Keywords: cognitive architecture, humanoid robots, artificial cognition, embodied intelligence, iCub, experience-driven learning, associative memory, intrinsic motivation, social robotics, developmental robotics, neurorobotics, event segmentation