I have spent a decade studying how human minds grow, and now I watch machines learn with a strange mixture of awe and discomfort. The advances in deep learning are remarkable, yet the gap between artificial systems and human learners seems to widen precisely when we try to make machines behave like children. In this article, I want to share a framework that I believe can bridge that gap: a cognitive growth pathway rooted in embodied intelligence. My central claim is straightforward. If we want artificial systems to develop flexible, transferable, and interpretable cognition, we must stop forcing them to learn from static datasets and instead let them grow through situated, embodied interaction with the world. I will argue that this pathway not only mirrors the earliest stages of human cognitive development, but also provides a practical blueprint for building embodied robots that learn the way infants do — through perception, action, feedback, and gradual abstraction.
## The Problem With Static Cognitive Models
Modern artificial intelligence systems have achieved superhuman performance in image classification, language translation, and game playing. Yet these same systems often fail when asked to reason about cause and effect, to adapt to a novel environment, or to build knowledge incrementally. The dominant paradigm — training deep neural networks on millions of labeled examples — treats intelligence as a process of pattern matching. The machine learns to associate inputs with outputs, but it never understands why the association holds. This is not intelligence; it is sophisticated curve fitting.
In educational settings, the limitation becomes even more visible. A child does not learn geometry by seeing a million labeled triangles. She learns by handling physical objects, by drawing, by failing, and by adjusting her mental models. She learns through her body — through her hands, her eyes, her movements. Traditional machine learning has no equivalent of this process. It cannot simulate the gradual construction of concepts that developmental psychologists have documented for over a century. It cannot model the role of action in perception, or the role of feedback in conceptual change. This is why I believe the next step for both artificial intelligence and educational technology lies in embodied cognitive growth.
## Theoretical Foundations
### Cognitive Development as Interactive Construction
Classical developmental theory teaches us that human intelligence emerges through interaction with the environment. The infant’s mind is not a blank slate waiting to be filled, nor is it a preprogrammed module. It is a dynamic system that builds itself through action. In the first two years of life, the child experiences the world through what we now call a sensorimotor stage. During this period, the child has no language, no formal logic, and no capacity for hypothetical reasoning. Yet the child is already constructing a coherent understanding of physical reality.
One of the most striking achievements of this stage is object permanence. In the first few months, a toy that moves behind a screen seems to cease to exist. The infant does not search for it. But after months of looking, reaching, and grasping, the infant begins to understand that the toy still exists even when hidden. This is not a trivial perceptual skill; it is a cognitive construction. The child learns that objects are continuous entities that persist through time and space.
For embodied robots, this insight is revolutionary. A robot that lacks object permanence will fail to track objects that are temporarily occluded. It will not understand that the cup still exists behind the barrier. By implementing a developmental progression similar to that of an infant, we can give an embodied robot the ability to maintain an internal representation of the world even when the sensory stream is interrupted. This is not simply a memory trick; it is the foundation of causal reasoning.
Another milestone is causal understanding. The infant discovers that shaking a rattle produces noise, that pushing a ball makes it roll, and that dropping a spoon makes it fall. These are not isolated observations. The infant builds a mental model of cause and effect through active experimentation. This is exactly what many machine learning systems lack. A statistical model can learn correlations between features, but it cannot distinguish causation from mere correlation. An embodied robot that explores its environment — that pokes, pushes, and pulls — can begin to construct genuine causal models from its own sensorimotor experiences.
A third critical element is agency. The infant learns that its own actions have consequences. When the child raises her hand, the mobile above the crib moves. This sense of agency is not a philosophical abstraction; it is the root of intentional behavior. For an embodied robot, agency means being able to predict the outcomes of its own actions and using those predictions to guide future behavior. Research in developmental robotics has shown that giving a robot the freedom to explore, to make errors, and to observe the results of its actions is far more effective for building robust motor skills than programing a fixed set of movements.
### Embodied Intelligence and Affordances
Embodied cognition theory argues that mental processes are shaped by the body’s morphology, sensorimotor capacities, and interactions with the physical and social environment. The classical view of perception — that the brain builds a rich internal representation of the world from impoverished sensory inputs — misses the point. Perception is not passive reception. It is an active, exploratory process. We see with our eyes, but we also see with our hands and feet. The world is not a collection of objects to be represented; it is a field of possible actions.
The ecological approach to perception introduced the concept of affordances — opportunities for action offered by the environment relative to an organism’s abilities. A chair affords sitting for a human, but it affords hiding for a cat and a climbing surface for a toddler. Affordances are relational; they depend on the body. This idea has profound implications for embodied robots. A robot must learn to perceive objects in terms of what they afford for its own body. A cup affords grasping, pouring, and throwing — but only if the robot’s gripper is shaped appropriately and its motor system can coordinate the necessary movements.
Embodied intelligence is not merely a philosophical stance. It suggests that true learning requires a body. The body is not a passive vehicle that carries a brain around; it is the very medium through which intelligence grows. For an embodied robot, this means that we cannot expect a disembodied neural network to understand the world by processing words or pixels alone. We must connect learning algorithms to a body and let the body interact with the environment.
### Limitations of Current Machine Learning
Let me be explicit about why I am dissatisfied with the current state of machine learning. The dominant paradigm is supervised statistical learning. We collect a massive dataset, annotate it, and train a deep network to minimize a loss function. This works well for tasks where the input-output mapping is stable and the training distribution matches the deployment distribution. But it fails spectacularly in open, dynamic, and uncertain environments.
Symbolic artificial intelligence emerged from the belief that intelligence could be encoded in logical rules. It works in well-defined domains, but it cannot cope with the vagueness and variability of real-world perception. Connectionist models, particularly deep neural networks, can handle perception but lack the mechanisms for causal inference, intentional action, and incremental knowledge construction. What we need is a third way — a path that combines the strengths of both approaches while adding the missing dimension: embodied, developmental learning.
The cognitive growth pathway I propose is designed to address these limitations. It is not a rejection of deep learning. Rather, it is an expansion of the learning paradigm. Instead of learning from static data, the system learns from interaction. Instead of a single, monolithic training phase, the system learns through a sequence of stages that mirror human development. Instead of opaque functions, the system builds interpretable structures that can be traced, debugged, and improved.
## The Cognitive Growth Pathway
### Framework: Three Stages of Emergent Cognition
I see the cognitive growth pathway as consisting of three main stages, each building on the previous one. The first stage is the sensorimotor stage. Here, the embodied robot has no predefined concepts and no symbolic reasoning. It only has a set of sensors and actuators. The robot must learn to map its actions onto the sensory consequences they produce. This is the stage of perception-action mapping. Through random exploration, the robot begins to discover that moving its arm causes a predictable visual change. It learns to track objects, to reach for them, and eventually to grasp them. The core task is not to classify objects, but to coordinate action and perception.
The second stage is the goal-directed stage. Once the robot has a basic sensorimotor repertoire, it can begin to act intentionally. It can choose an action based on a desired outcome. This requires a primitive understanding of cause and effect. The robot must learn that if I push this object, it will fall; if I pull that handle, the drawer will open. This kind of causal learning is far more powerful than statistical pattern recognition because it allows the robot to plan novel action sequences in new situations. In this stage, the robot also begins to develop a sense of agency — it knows that its own actions can change the state of the world.
The third stage is the symbolic learning stage. Here, the robot starts to abstract from concrete sensorimotor experiences. It learns concepts such as “object,” “container,” “support,” and “above.” These concepts are not arbitrary labels; they are grounded in the robot’s own interactive history. The robot can now reason about objects it has never seen before, because it understands the relational structure of the world. This stage is analogous to the development of language and symbolic thought in children. It enables the robot to learn from instruction, to imagine hypothetical scenarios, and to generalize across contexts.
Table 1 summarizes the three stages and their corresponding machine learning concepts.
| Developmental Stage | Key Cognitive Achievements | Machine Learning Analogue |
|---|---|---|
| Sensorimotor | Perception–action mapping, object permanence, body schema | Self-supervised learning, reinforcement learning with low-level motor control |
| Goal-directed | Causal inference, agency, intentional planning | Model-based reinforcement learning, inverse dynamics models, intrinsic motivation |
| Symbolic | Concept formation, abstraction, relational generalization | Symbol grounding, compositional representation learning, meta-learning |
### Mechanisms Underlying Cognitive Growth
The transition from one stage to the next is not automatic. It requires at least three mechanisms: embodied regulation, situated construction, and representational reorganization.
Embodied regulation refers to the way the body continuously calibrates the learning process. The robot’s own morphology — the length of its limbs, the arrangement of its sensors, the force limits of its actuators — determines what actions are possible and how sensory data should be interpreted. This is not a nuisance to be overcome; it is a powerful source of inductive bias. An embodied robot that has two arms and a head with two cameras will naturally learn different concepts than a wheeled platform with a single laser scanner. Body structure shapes cognition.
Situated construction is the process by which knowledge is built in the context of specific situations. A robot does not learn about cups in the abstract. It learns about cups by encountering them on a particular table, in a particular light, with a particular gripper. The challenge is to extract general principles from these specific encounters. This is where the notion of “decoupling” becomes useful. The robot must learn to separate the invariant properties of a cup from the variable features of the situation. In machine learning, this is analogous to learning disentangled representations. But for an embodied robot, the separation is guided by action: the invariant properties are those that persist under different actions and viewpoints.
Representational reorganization is the process of restructuring internal models as new experience accumulates. This is not simply adding new memories. It involves changing the categories and relational structures that organize existing knowledge. For example, a robot that has learned to grasp cups and bottles may later come to understand the more abstract category “drinkable container.” This requires a reorganization of its internal representation. In neural networks, this can be achieved through curriculum learning, where the difficulty of training examples is gradually increased, allowing the network to form successively higher-level features.
These mechanisms are not independent. They work together in a continuous cycle. The body regulates the learning process, experience constructs situated knowledge, and the system reorganizes its representations to accommodate new insights.
### Computability of Cognitive Growth
If the cognitive growth pathway is to be more than a metaphor, it must be realized in computational systems. I have found that the developmental learning process can be formalized as a Markov Decision Process. In an MDP, the embodied robot is an agent that interacts with its environment through a sequence of states, actions, and rewards. The agent’s goal is to learn a policy that maximizes cumulative reward. The MDP is defined by the tuple \((\mathcal{S}, \mathcal{A}, \mathcal{T}, \mathcal{R}, \gamma)\), where \(\mathcal{S}\) is the set of states, \(\mathcal{A}\) the set of actions, \(\mathcal{T}(s’|s,a)\) the transition probability, \(\mathcal{R}(s,a,s’)\) the reward function, and \(\gamma \in [0,1]\) the discount factor.
The optimal action-value function \(Q^*(s,a)\) satisfies the Bellman equation:
$$
Q^*(s,a) = \mathbb{E}\left[ \mathcal{R}(s,a,s’) + \gamma \max_{a’} Q^*(s’,a’) \right]
$$
This equation elegantly captures the essence of goal-directed learning: the value of an action is the sum of the immediate reward and the future rewards that follow from the resulting state. For an embodied robot, this is exactly how causal understanding is built. The robot learns to choose actions that lead to desired outcomes because the Bellman equation relates immediate actions to long-term consequences.
Reinforcement learning provides the algorithmic machinery for optimizing the policy. Q-learning updates the action-value estimate using the rule:
$$
Q(s,a) \leftarrow Q(s,a) + \alpha \left[ r + \gamma \max_{a’} Q(s’,a’) – Q(s,a) \right]
$$
where \(\alpha\) is the learning rate. This update is fully compatible with sensorimotor contingencies. The robot experiences a state, executes an action, observes a reward and the next state, and updates its value estimates. Over time, the robot develops a predictive model of its environment, which is the computational equivalent of an infant’s understanding of object permanence and cause-effect relations.
Self-supervised learning is equally important. In self-supervised learning, the robot generates its own training signals from raw sensory data. For example, an embodied robot can learn visual representations by predicting the next frame in a video stream, or by predicting the effect of its own action on the image. This is exactly what a baby does when she moves her hand and watches it change shape in her visual field. The robot does not need external labels; it uses the structure of the sensorimotor loop as its own teacher.
Attention mechanisms and compression strategies further support cognitive growth. The robot must select from an overwhelming stream of sensory data the features that are most relevant to its current task. This is similar to the action of attention in human perception. By attending to focused regions of the visual field and encoding compact representations, the robot reduces the computational cost of learning and improves generalization. A simple way to express the compression process is through an information bottleneck objective:
$$
\min_{p(z|x)} \; \mathcal{L} = – \mathbb{E}_{z \sim p(z|x)} \left[ \log q_{\phi}(x|z) \right] + \beta \, D_{\mathrm{KL}}\left( p(z|x) \| r(z) \right)
$$
Here, \(x\) is the sensory input, \(z\) is a compressed latent representation, \(q_{\phi}\) is a decoder, and \(r(z)\) is a prior. The first term encourages accurate reconstruction, while the second term encourages information compression. This kind of variational autoencoding can serve as a mechanism for representational reorganization in an embodied robot.
Figure 1 in my original work shows the computational framework connecting these ideas. The cognitive growth demand is translated into a Markov decision process for perception-action decision modeling, reinforced by RL and self-supervised learning for autonomous optimization and pattern discovery, and refined by attention mechanisms and compressed representations. The result is machine cognitive growth — active learning in an embodied interactive context.
## Comparison With Traditional Machine Learning
To appreciate the significance of the cognitive growth pathway, I place it side by side with traditional machine learning in a systematic comparison.
| Aspect | Traditional Machine Learning | Cognitive Growth Pathway |
|---|---|---|
| Learning paradigm | Batch training on static datasets | Online, incremental, interactive learning |
| Data requirement | Massive labeled data | Fewer labels, using self-generated signals |
| Role of body | Irrelevant or ignored | Central to perception and action |
| Causal reasoning | Limited to correlational patterns | Constructed through goal-directed exploration |
| Generalization | Fragile under distribution shift | Robust through grounded concepts |
| Interpretability | Opaque or black-box | Structured, stage-wise, and traceable |
| Adaptation to novelty | Often requires retraining | Continuously adapts and reorganizes |
The contrast is stark. Traditional machine learning excels at pattern matching but struggles with understanding. The cognitive growth pathway does not reject the power of deep networks; rather, it embeds them within a developmental architecture that gives them purpose, structure, and meaning. The result is not just a more efficient learner, but a more human-like learner.
## Implications for Education
I have moved between artificial intelligence and education for a long time, and I am convinced that the cognitive growth pathway can transform how we think about teaching and learning. The pathway provides a computational metaphor for educational processes that have been described qualitatively for decades. More importantly, it offers a methodology for constructing “virtual learners” — synthetic agents that can simulate how real students learn and develop under different conditions.
These virtual learners are not simple chatbots. They are embodied robots embedded in simulated environments, complete with sensors, actuators, and a developmental curriculum. They can be used to test educational theories at scale. Suppose we want to know whether discovery-based learning or direct instruction is more effective for a particular concept. We can create two virtual learners — one that learns through free exploration, and another that receives explicit training — and compare their cognitive growth trajectories. This goes far beyond static statistical analysis; it allows us to watch the process of knowledge construction unfold in real time.
The pathway also emphasizes intrinsic motivation. In the goal-directed stage, the embodied robot is not driven by external rewards alone; it is motivated by curiosity, by the desire to explore novel situations, and by the satisfaction of gaining control over the environment. This is exactly what educational psychologists call “autonomous learning motivation.” By implementing self-purposed reinforcement learning frameworks, we can model how students set their own goals and pursue them without external incentives. The insights from these models can inform classroom design: create environments that offer novelty, allow choice, and support gradual mastery.
Another connection is to the concept of “learning by doing.” The cognitive growth pathway shows that knowledge is not transmitted from teacher to student; it is constructed through action and feedback. This is the computational equivalent of the famous educational principle that we learn by doing. An embodied robot that learns to grasp by trying 50,000 times is demonstrating the value of hands-on experience. For human students, this suggests that abstract concepts become meaningful when they are grounded in physical or simulated action.
I also see applications in cognitive development assessment. Traditional tests measure the product of learning — what the student knows at a given moment. The cognitive growth pathway allows us to measure the process: how the student acquires knowledge, how they debug errors, and how they reorganize their mental models. By embedding a virtual learner in a standardized environment, we can observe learning strategies and predict long-term outcomes. This is especially promising for children with developmental differences, as the parameters of the virtual learner can be adjusted to simulate different cognitive profiles.
Teacher professional learning is another area that could benefit. Teachers often struggle to understand the difference between a student who memorizes a procedure and one who genuinely understands the underlying concept. The cognitive growth pathway provides a transparent model of conceptual change that can illustrate this difference. By interacting with an embodied robot that learns the same topic, teachers can see how understanding emerges from action and reflection. This can be used in teacher training to demonstrate why constructivist pedagogy works and why rote drill often fails.
## Advantages and Limitations
The cognitive growth pathway offers several distinct advantages over current approaches. First, it is biologically plausible. It builds on the well-documented stages of human development, giving us confidence that the architecture is not arbitrary. Second, it is interpretable. Because the learning process is segmented into stages and driven by explicit mechanisms, we can always ask why an embodied robot made a particular decision. Third, it is resilient. By continuously interacting with the environment, the robot learns to handle changes and adapt to new situations without full retraining. Fourth, it reduces the need for labeled data. The robot learns from its own experience, which is especially valuable in domains where labels are scarce or expensive.
However, the path is not without obstacles. The most obvious challenge is computational cost. Developmental learning requires the agent to explore the environment for thousands of interactions, which can be prohibitively expensive in both simulation and real hardware. This is not just a matter of compute; it is also a matter of time. In contrast, a supervised deep network can be trained in hours or days on a static dataset. An embodied robot might need days of real-world interaction to learn a single motor skill. In simulation, this is more feasible, but the gap between simulation and reality remains a major challenge.
Another limitation is the incomplete mapping between psychological mechanisms and neural implementation. We know that infants engage in sensorimotor play, but we do not fully understand how the brain translates these experiences into concepts. Our models may capture the functional aspects of cognitive growth, but they fail to capture the rich biological substrate — hormones, emotion, social interaction, and neural plasticity — that accompanies human development. For embodied robots, this means that we are building a thin caricature of human cognition, not a faithful replica.
A further concern is the difficulty of evaluating progress. When a deep learning model achieves a certain accuracy, we know exactly what it can do. When an embodied robot has gone through a developmental trajectory, it is less clear what it has “learned.” We need new evaluation metrics that capture the structure and flexibility of the acquired knowledge, not just task performance. This is an open research problem that will require collaboration between educators, psychologists, and roboticists.
## Moving Forward
Despite these challenges, I am optimistic about the future of embodied cognitive growth. The convergence of developmental robotics, machine learning, and educational psychology is creating a new scientific field that promises to reshape both artificial intelligence and education. I envision a future where every classroom has a virtual embodied robot that can model the thinking process of each student. The teacher can ask the robot: “Why did you make this mistake?” and the robot can reply with a clear explanation based on its developmental history. I imagine educational tools that adapt not just to the student’s performance level, but to the student’s cognitive strategy, using the cognitive growth pathway as the underlying engine.
For embodied robots, the path forward is clear. We must build systems that can explore, act, and receive feedback in real environments. We must give them bodies that are rich enough to support the development of concepts. We must allow them to grow slowly, stage by stage, just as human children do. And we must be patient — true intelligence cannot be rushed.
I have written this article as a reflection on my own research journey, but also as a call to fellow researchers and educators. The cognitive growth pathway is not a definitive answer; it is a direction. It is a commitment to viewing intelligence as a dynamic, embodied, and developmental phenomenon. It is a rejection of the static, disembodied, and brittle models that dominate current artificial intelligence. And it is an invitation to think about learning in a new way — not as the transfer of information, but as the growth of a mind through interaction with the world.
In the end, the goal is not to build machines that mimic human performance on tests. The goal is to build embodied robots that can learn the way we do: by playing, by wondering, by making mistakes, and by gradually coming to understand the beautiful complexity of the world. This is the promise of the cognitive growth pathway, and I believe it will guide us toward a richer, more humane form of intelligence — both artificial and natural.