Developmental robotics

• 1. Introduction

  ›

  Developmental robotics asks a question that would have seemed absurd not long ago: can a machine learn the way a child does? Not by being programmed with every rule, but by exploring, making mistakes, and gradually building up knowledge from scratch. The field, sometimes called epigenetic robotics, wants robots that discover their own bodies, pick up language, develop social skills, and keep learning across an entire lifetime, without an engineer stepping in to update their instructions.

  Alan Turing and other pioneers of cybernetics already sketched out this vision as early as 1950. But it took until the end of the twentieth century for researchers to begin investigating those questions in a sustained, systematic way. Since then, a community has formed at the intersection of robotics, developmental psychology, neuroscience, evolutionary biology, and linguistics, each field feeding ideas into the others.

  The goal is not a robot that can do any one task brilliantly. The goal is a robot whose set of skills can keep growing, potentially without limit, the way a child's can. What mechanisms make that possible? What happens when those mechanisms break down? And what can robots teach us about how human infants actually learn?

• 2. Child as Blueprint

  ›

  Developmental psychology, neuroscience, and evolutionary biology all sit at the foundation of the developmental robotics method. Researchers begin with verbal or descriptive theories about how children and animals develop, then formalize those theories into computational models and implement them in physical robots. The robot becomes a kind of test bench: if the model is correct, the robot should behave in ways the theory predicts.

  That process runs in both directions. When a model fails in the real world, it forces researchers to revise their biological assumptions. Developmental robotics therefore does not just borrow from the life sciences; it sends hypotheses back to them. A finding about how a simulated robot fails to acquire a skill can suggest new experiments for a developmental psychologist studying infants.

  Linguistics also plays a significant role, because language acquisition is one of the central puzzles researchers want to crack. The field uses the term "symbol grounding" to describe the problem of connecting words and abstract symbols to the sensorimotor experiences that give them meaning, a challenge that applies equally to children and robots learning their first words.

• 3. Skill by Skill

  ›

  The skills that developmental robots are taught to acquire mirror, in sequence and structure, the skills that human infants develop. Sensorimotor skills come first: discovering the structure and dynamics of one's own body, learning hand-eye coordination, locomotion, and how to interact with and use objects. A particular focus lands on affordances, which means learning what actions a given object or environment makes possible.

  Social and linguistic skills form a second major category. Researchers investigate how robots can learn turn-taking, engage in coordinated interaction, and build up a working lexicon and eventually syntax and grammar. Crucially, those linguistic skills must be grounded in sensorimotor experience rather than floating free as pure symbol manipulation.

  Running alongside both of those is a third layer: cognitive skills. These include the emergence of a self/non-self distinction, the development of attention, the formation of categories, and higher-level representations of affordances and social constructs. Researchers also look at how values, empathy, and something resembling a theory of mind might develop in an artificial system.

• 4. What Drives Exploration

  ›

  The sensorimotor and social spaces a robot inhabits are so vast that no system could ever explore all of them in a lifetime. Mechanisms and constraints are therefore essential to guide development and keep the growth of complexity under control. Developmental robotics has identified several families of these guides, each drawn from observations of biological development.

  Motivational systems generate internal reward signals that steer exploration. Extrinsic motivations push a robot toward maintaining basic properties such as physical integrity or energy balance, roughly analogous to hunger or pain in animals. Intrinsic motivations work differently: they push the robot toward novelty, challenge, or what researchers call learning progress for its own sake. This kind of drive is sometimes described as curiosity-driven or active learning.

  Social guidance is a separate but related mechanism. Because humans learn enormous amounts through interaction with peers, developmental robotics examines how robots can read and respond to social cues in ways that allow humans to teach them naturally, through demonstration, imitation, and other forms of everyday pedagogy. The field also investigates statistical inference biases that make learning more efficient, and mechanisms that let a robot reuse previously acquired knowledge when building new skills.

  Embodiment itself acts as a constraint. The geometry and materials of a robot body, and any built-in motor primitives encoded as dynamical systems, can dramatically simplify the acquisition of skills. Researchers refer to this as morphological computation. Maturational constraints round out the picture: just as the human body and nervous system grow progressively after birth, a robot's available sensorimotor signals and degrees of freedom can be allowed to expand over time, and studying how that progression helps or hinders learning is a central research question.

• 5. A Field Finding Its Limits

  ›

  No experiments in developmental robotics have yet run longer than a few days. That gap matters because human infants, equipped with brains and bodies far more powerful than anything robotics currently offers, still need years to acquire basic sensorimotor skills. The contrast makes clear how far current systems are from the field's own goals.

  High-dimensional continuous sensorimotor spaces are a recognized technical obstacle. Lifelong cumulative learning, the capacity to keep adding genuinely new skills rather than overwriting old ones, has not been solved. And the mechanisms studied so far have largely been investigated in isolation from one another. What happens when intrinsically motivated learning, socially guided learning, and maturational constraints all operate simultaneously and interact is still an open question.

  Human-robot interaction poses its own challenge. Robots currently struggle to perceive and interpret the full diversity of social cues that non-engineer humans provide in natural settings. That limitation blocks the kind of everyday teaching that would allow a robot to learn from ordinary people rather than from specialists.

  One of the deepest unsolved problems cuts across both robotics and biology: how do compositionality, functional hierarchies, and modular structures emerge from development? This connects directly to the symbol grounding problem in language, and to contested questions about whether symbolic representations in the brain are even necessary, or whether alternative frameworks allowing for compositionality might replace them. The field labels this area active and unresolved.

• 6. Evolution and Development Together

  ›

  During biological epigenesis, a body's morphology does not arrive fixed at birth. It develops in continuous interaction with the sensorimotor and social skills the organism is acquiring. Translating that into robotics creates obvious engineering challenges, but researchers argue it may be a crucial mechanism rather than an optional one. Morphogenetic robotics, which explores developing morphologies at least in simulation, is one avenue being pursued.

  In biology, developmental mechanisms operating at the scale of a single lifetime interact closely with evolutionary mechanisms operating across generations. The growing "evo-devo" literature documents how tightly those two timescales are intertwined. In artificial organisms, including developmental robots, that interaction remains vastly understudied.

  The NSF and DARPA funded a Workshop on Development and Learning held on the 5th-the 7th of April 2000 at Michigan State University. It was the first international meeting devoted to the computational understanding of mental development by robots and animals. The organizers chose the phrasing "by robots and animals" deliberately: the agents were understood to be active participants in their own development, not passive subjects. From that 2000 gathering, two distinct conferences eventually merged into a single annual event, the ICDL-EpiRob conference, which has brought the community together since 2011.