Latest / Embodied AI 101 / The Crib as a Causal Laboratory: How a Playing iCub Learns "I Did That"
Transcript
- The Crib as a Causal Laboratory: How a Playing iCub Learns “I Did That”. Robot in a crib: How a playing robot helps us understand sensorimotor contingency learning • Josua Spisak, Sergiu Tcaci Popescu, Lukáš Rustler, Stefan Wermter, J. Kevin O’Regan, and Matej Hoffmann • University of Hamburg, Czech Technical University in Prague, Université Paris Cité, and CNRS • Science Robotics, volume 11, issue 117 • 2026. Picture an iCub humanoid lying on its back in a crib. A mobile hangs above it. The robot moves, the mobile responds, and the controller tries to work out a deceptively difficult fact: which changes in the sensory world are consequences of its own actions?. That scene could easily be dismissed as a charming developmental-robotics demonstration. But the scientific question underneath it is fundamental. Before an agent can manipulate an object, compensate for damage, use a tool, or attribute an event to another actor, it must discover the causal structure connecting motor commands to sensory consequences. It needs to learn, in some operational sense, “when I do this, that happens.”. Published on August 26, 2026—three days before the date of this episode—Robot in a crib argues that this learning process is not adequately summarized by a familiar behavioral statistic such as how much a limb moves. The robot successfully detects sensorimotor contingencies, but it does so through multiple strategies emerging from a changing balance between prediction and curiosity. The interesting result is therefore not merely that a humanoid can reproduce a classic infant experiment. It is that apparently different behaviors may express the same underlying causal competence. The smallest possible causal laboratory. The developmental-psychology setup behind this work is known as the mobile paradigm. In its classic form, a ribbon connects one of an infant’s limbs to a mobile suspended over a crib. Movement of the connected limb produces movement in the mobile. Researchers then ask whether the infant discovers this relationship and changes its behavior accordingly. The paradigm dates to work by Carolyn and Michael Rovee in the late 1960s and by John Watson and Suzanne Ramey in the early 1970s. Studies have reported that infants as young as roughly two months increase their overall activity when their movement controls the mobile. By five or six months, infants can more reliably differentiate the connected limb, increasing its activity relative to the others. Crucially, experiments can compare contingent stimulation with a control condition in which the mobile moves just as often but independently of the infant’s actions. That comparison separates sensitivity to controllability from a simpler preference for visual stimulation. This is a remarkably compact causal-learning task. The agent has several possible actions, a salient sensory event, and a hidden dependency between one action channel and that event. There is no need for language, instruction, symbolic goals, or externally specified task reward. The experiment also contains useful variations. In a binary condition, crossing an activity threshold can trigger a fixed mobile response. In a conjugate, or non-binary, condition, mobile movement varies proportionally with limb activity. Researchers can also remove the connection after learning. Infants sometimes respond with an “extinction burst”: a short-lived increase in movement, as though testing a prediction that has suddenly stopped working. Neither the extinction burst nor every reported contingency effect is universally replicated, however. Recent re-examinations have emphasized that the empirical literature is less clean than its textbook summary suggests. That uncertainty makes the mobile paradigm more, rather than less, interesting for computational modeling. If every infant displayed the same monotonic rise in kicking, a simple reinforcement rule might be sufficient. Variable trajectories force us to ask what internal mechanisms could generate several outwardly different but still sensible responses. Why “more movement” is not equivalent to learning. Suppose an infant moves the connected leg more frequently. At least three broad explanations are available. One is operant reinforcement: mobile movement is stimulating, so actions producing it become more probable. A second is largely physical: coupling the limb and mobile creates a dynamical system whose mechanical feedback reorganizes movement. A third is predictive and exploratory: the infant learns an action–effect model, intentionally or implicitly probes it, and changes its behavior as uncertainty falls. All three can produce increased connected-limb activity. Looking only at the final movement count therefore creates a model-identification problem. The same behavioral curve may be compatible with very different mechanisms. The inverse problem is equally important. A learner that has detected the contingency need not keep increasing its movement. It might make a few diagnostic motions and then stop. It might find an efficient rhythm that moves the mobile with little effort. It might hold a limb in a favorable region and use sparse pulses. Or it may alternate action and stillness to distinguish self-produced motion from the mobile’s residual oscillation. The central wager of Robot in a crib is that these differences are not necessarily noise around one canonical learning response. They can be signatures of distinct exploratory strategies. What the new paper contributes. The published study implements an embodied computational learner on a child-sized iCub interacting with a real mobile. Its learning is organized around two mechanisms: predicting the sensory consequences of action and remaining curious about action regimes whose consequences are not yet well understood. Across multiple experimental conditions, the robot detects the relevant contingencies while producing a range of movement strategies. Analysis of its internal activity ties that behavioral diversity to the changing balance between prediction and curiosity—or, in familiar machine-learning language, exploitation and exploration. The study grows out of an earlier computational manuscript by Spisak and colleagues, first posted in April 2025. That precursor modeled the mobile paradigm in simulation and targeted several established observations: selective activity of the connected limb, differences between contingent and non-contingent stimulation, binary versus conjugate feedback, and occasional extinction bursts. The Science Robotics paper adds Lukáš Rustler to the author team and makes physical embodiment central to the investigation. This distinction matters when reading the technical lineage. The earlier manuscript exposes the architecture in considerable detail, while the journal paper’s headline result concerns what happens when prediction and curiosity encounter a real robot, a real mobile, and real dynamics. The precise numerical architecture of the precursor should not automatically be assumed unchanged in every detail, but it provides the clearest account of the computational idea being embodied. Under the hood: prediction, curiosity, and preferred activity. The precursor model is not conventional reinforcement learning. There is no externally defined reward and no objective to maximize discounted return. Instead, the learner issues motor activity, predicts the mobile’s response, observes what actually happens, and uses the mismatch to reorganize future exploration. The authors explicitly distinguish this process from reward maximization. Prediction is more than a diagnostic output. The controller learns a forward relationship between its activity and expected sensory feedback. Prediction error is therefore actionable. If the robot performs a movement and the mobile behaves differently from what was expected, that part of action space becomes interesting. In the precursor, a neural controller produced a high-dimensional set of muscle-like commands together with a prediction of the next sensory effect. A many-to-one mapping transformed 600 motor-command outputs into activity across four limbs, with motor noise adding variability. The complete output contained 601 units: 600 associated with motor commands and one associated with sensory prediction. This deliberately redundant motor representation approximated the fact that biological movement is not selected from four clean buttons labeled “move left arm,” “move right arm,” and so forth. That redundancy is scientifically consequential. If many internal commands can produce similar macroscopic movement, the learner must discover useful regions of a motor manifold rather than merely identify one privileged discrete action. It also creates room for several strategies to arrive at similar sensory effects. Curiosity is an action scheduler, not a bonus prize. The precursor’s exploration mechanism maintains an activity-interest map. For each limb, activity is divided into ranges. Each range has an interest value. A movement producing an unexpected sensory outcome raises the interest associated with that range; a correctly predicted outcome reduces it. In the worked example, surprise pushes an activity range’s interest upward, while a predictable result causes interest to decay and allows another range to be sampled. This is subtler than simply maximizing prediction error. The controller does not receive one global scalar bonus saying that surprise is good. Curiosity is indexed by body part and movement magnitude. It changes where the learner samples next. The resulting loop is competence-sensitive:. The system tries a movement. An unexpected outcome makes that movement regime worth revisiting. Repeated experience improves the prediction. Once the outcome becomes unsurprising, interest falls. Exploration migrates toward another unresolved region. This mechanism can generate non-monotonic behavior without invoking fatigue or a hand-authored phase schedule. Early surprise can increase activity, successful learning can reduce it, and a later violation can revive exploration. Homeostasis prevents curiosity from becoming the entire organism. The model also includes a preferred, or baseline, activity level. Its learning objective combines pressure to predict sensory feedback, pressure to execute movements selected by the interest mechanism, and pressure not to deviate arbitrarily far from baseline activity. In plain language, the learner wants to understand what its actions do, investigate unresolved possibilities, and remain within an endogenous range of motor behavior. That baseline term is easy to overlook, but it is relevant to autonomous robotics. An agent with no external task should not necessarily become motionless, nor should it maximize movement indefinitely. Developmental systems need some internally generated level of activity from which exploration can begin. Ablations in the precursor suggested that action–outcome prediction, curiosity-driven exploration, motor noise, and a sufficiently rich motor representation were important for reproducing core mobile-paradigm behavior. Reducing the number of muscle-like commands or removing exploration changed the resulting trajectories. By contrast, a particular mechanism intended to alter baseline “fussiness” over time was judged unnecessary. Why putting the model in a body changes the scientific question. A simulation can specify that a limb activity of 0.6 causes a sensory response of 0.6. A physical mobile cannot be reduced so easily. The mobile has inertia. Its current motion depends on previous interactions. A command may have a delayed effect, and the same command can produce different sensory results depending on pose, velocity, ribbon tension, oscillation phase, actuator dynamics, and sensor latency. What looks like noise to a memoryless learner may actually be hidden state in the environment. This makes the real setup a partially observed system. To discover the contingency, the robot may need not only to select a limb but also to place that limb appropriately, time the motion, wait for residual mobile movement to decay, or exploit the mobile’s existing momentum. The authors’ own description of the emergent strategies emphasizes spatial positions, environmental dynamics, and particular movements—not merely total activity. Physical embodiment therefore turns the experiment from simple correlation detection into active system identification. The robot chooses interventions while simultaneously learning a model from the data those interventions generate. Its exploration policy affects the observations available to its predictor; the predictor’s errors then reshape exploration. This feedback is central to embodied learning. The learner does not passively receive a fixed dataset sampled from the world. It creates its training distribution through movement. The binary-versus-conjugate distinction makes that point concrete. Binary feedback offers high causal contrast: either the threshold was crossed and the mobile responded, or it did not. Conjugate feedback offers a richer relationship between action magnitude and outcome, but that relationship is more exposed to physical variation. In the precursor simulations, connected-limb differentiation was clearer under binary feedback, while aggregate evidence for an extinction burst appeared more strongly in the non-binary case. One interpretation is that binary feedback is easier to detect, whereas continuous feedback supports more nuanced control once discovered. Those are different learning problems, and they need not yield the same outward behavior. The main result: learning is a family of strategies. The paper’s most important claim is that successful contingency learning cannot be reduced to one behavioral metric. The robot detects action–effect relationships, yet different runs express that competence differently. The activity of the connected limb remains informative, but it is not a sufficient statistic for the learning process. Consider two failure modes for a movement-count metric. The first is a false negative. A robot may learn that a short pulse at a particular pose reliably excites the mobile. Once that relation is predictable, repeatedly making large movements is unnecessary. Average activity can decline even as control improves. The second is a false positive. A robot—or infant—can move vigorously because the mobile is visually exciting, because the body is generally aroused, or because the system is chasing unpredictable stimulation. High movement does not by itself establish that an action-specific contingency has been learned. Trajectory structure carries the missing information. Relevant features can include which limb moves, where it is positioned, the amplitude and frequency of motion, the ordering of movements, the lag between action and sensory response, and the use of pauses. A pulse–wait–pulse sequence, for example, can be more causally diagnostic than continuous kicking because stillness allows the learner to see whether the environment continues changing without further intervention. The prediction–curiosity loop provides a plausible generator for this diversity. When a region of action space is poorly predicted, curiosity pulls the controller toward it. As the predictor improves, that pull weakens. Meanwhile, predictable actions can be exploited to produce desired or informative effects. Because prediction quality differs across body configurations and across runs, each learner can travel through the action space in a different order. Here, exploitation should not be read as reward harvesting. There is no task reward to harvest. It means relying on a sensorimotor relationship that has become predictable or controllable. Curiosity pulls toward unresolved relations; prediction supports economical use of known ones. This also offers a restrained account of extinction bursts. When the mobile is disconnected, an established prediction suddenly fails. The resulting surprise can make the formerly effective movement interesting again, producing a temporary burst of testing. But if a run never established a strong prediction, had already shifted elsewhere in action space, or used a strategy not based on high-amplitude repetition, no visible burst need occur. The precursor simulations indeed showed substantial run-to-run variability, with extinction bursts in some trajectories but not others. The broader lesson is methodological. Averaging across runs may erase the mechanism one hopes to study. If half the learners adopt vigorous rhythmic excitation and the other half adopt sparse diagnostic probes, the mean trajectory may resemble neither strategy. The repository accompanying the journal article explicitly mentions movement-analysis images and decision-tree cluster-analysis artifacts. That is consistent with treating behavioral variation as structured data to be interpreted rather than variance to be discarded. Where this sits among neighboring theories. The mobile paradigm has inspired several quite different computational explanations. Work by John Watson, Nicholas Butko, and Javier Movellan framed contingency detection in information-theoretic terms. Their agents selected behavior to reduce uncertainty and identify responsive partners, motivated partly by infants’ social reactions during contingent interactions. This line is close to active information gathering, although the original authors acknowledged that exact information-maximizing calculations may be biologically demanding. Zaadnoordijk and colleagues’ Babybot simulations demonstrated that simple operant reinforcement can reproduce the core rise in connected-limb activity. Their intentionally minimalist controller repeated actions that produced stimulation and did not represent or predict the action–effect relationship. This is valuable because it establishes a lower bound: one should not invoke rich causal cognition merely to explain increased kicking. But the absence of prediction makes expectancy violations and extinction bursts harder to explain within the same mechanism. Kelso and Fuchs instead modeled the infant–mobile setup as a coupled dynamical system. On that account, behavioral transitions can emerge from the coordination dynamics of the body, ribbon, and mobile without reward, prediction, or an explicit representation of agency. Related dynamical-systems work has examined how physical agency and differentiated action can emerge through the mechanics of coupling. Robot in a crib occupies a productive middle ground. It takes physical dynamics seriously but also assigns a functional role to internal prediction. It accepts that simple reinforcement can generate one behavioral signature, while asking whether curiosity and forward prediction can explain a wider distribution of trajectories. There is an obvious family resemblance to predictive processing, active inference, intrinsic motivation, empowerment, and modern world-model agents. But the implementation should not be collapsed into any one of those labels. It does not derive every component from a single free-energy objective, and it is not a standard RL agent with prediction error appended as an intrinsic reward. Its distinctive choice is to make curiosity local to particular sensorimotor regimes and let that structured interest map participate directly in action selection. For contemporary robotics, that is the useful abstraction: learn a forward model, track where that model remains inadequate, and direct physical exploration toward those regions while respecting endogenous activity constraints. Hypotheses this generates for infant research. A computational model earns its place in developmental science not by looking infant-like, but by generating discriminating predictions. First, successful learners should not always move more. Some infants may show an early burst of probing followed by efficient or reduced activity once the contingency becomes predictable. Classifying learning solely through mean movement risks excluding precisely those infants who use economical strategies. Second, temporal organization should outperform raw activity counts. Measures of action–outcome delay, pause structure, rhythmicity, pose occupancy, and changes in movement variability may separate contingent learners from stimulation-seeking controls more effectively than total movement alone. This follows directly from the robot’s use of environmental dynamics and specific movement configurations. Third, extinction bursts should depend on the learner’s internal history. A burst should be most likely when the learner has acquired a confident prediction, continues to assign interest to the relevant movement, and then experiences an abrupt violation. It should not be treated as a mandatory marker of agency. Fourth, binary and conjugate feedback should alter the distribution of strategies, not merely the speed of learning. Binary feedback may encourage threshold-crossing actions. Continuous feedback may favor amplitude tuning, rhythm, and exploitation of the mobile’s dynamics. Two groups could therefore learn equally well while producing different kinematic signatures. Fifth, non-contingent stimulation may create global surprise without body-part specificity. A randomly moving mobile is interesting, but no particular limb regime should become reliably predictive. Internal curiosity could remain elevated while the connected-versus-unconnected behavioral contrast fails to emerge. These hypotheses are experimentally accessible. Modern motion capture, inertial sensing, and time-series analysis make it possible to retain the individual trajectories that older analyses often compressed into minute-level activity bins. What autonomous-robotics researchers should borrow. The immediate value of this work is not that industrial or household robots should be placed in cribs. It is that the mobile paradigm distills a design pattern for autonomous self-calibration. A robot receiving a new end effector must discover which commands move it and which visual or tactile changes are self-produced. A damaged robot must identify which portions of its old sensorimotor model remain valid. A tool-using robot must learn how motions of the body propagate through an attached object. An active-perception system must distinguish scene motion from camera-induced motion. In each case, the agent needs structured exploration before a downstream task reward is available. Several engineering lessons follow:. Separate controllability from stimulation. Include matched controls in which sensory events occur with the same frequency and magnitude but are not action-contingent. Index curiosity by action context. A single global novelty score cannot tell the controller which joint, pose, force range, or contact mode requires investigation. Give the agent endogenous operating constraints. Preferred activity, energy limits, safety envelopes, and recovery behavior should shape exploration from the beginning. Preserve individual trajectories. Evaluate strategy distributions, switching events, pauses, timing, and internal prediction quality rather than reporting only average action magnitude. Learn inside the real dynamics. Simulation remains useful, but latency, compliance, contact, inertia, and motor redundancy can change what constitutes an informative action. For modern world-model architectures, the natural extension would be a recurrent and uncertainty-aware predictor. Rather than driving curiosity with raw error, one could estimate epistemic uncertainty or learning progress, reducing the familiar “noisy television” failure in which irreducibly random outcomes remain perpetually attractive. Curiosity could also operate hierarchically: first over body regions, then motor primitives, and finally temporally extended skills. The model could be paired with foundation-model representations without surrendering its developmental logic. A pretrained visual encoder might describe the environment, while the robot still has to discover from first-person intervention which latent changes it can cause. Semantic knowledge that an object is a “mobile” is not the same as embodied knowledge of how this body moves that mobile. There is also a practical reproducibility caveat. The authors’ Zenodo record describes about 1.4 gigabytes of Python-related resources, condition-specific datasets, ablations, movement imagery, and cluster analyses. As of August 29, 2026, however, the public record marked the files as restricted, and the linked GitHub repository did not resolve through a public request. Researchers planning immediate replication may need to check again or contact the authors. What the experiment does not prove. A robot model can demonstrate computational sufficiency: prediction plus structured curiosity can generate contingency-sensitive behavior and diverse exploratory strategies. It cannot establish that infants use exactly the same algorithm or neural representation. The iCub also arrives with substantial engineered structure. Its body parts, sensors, actuator interfaces, and learning channels have already been defined by designers. The system learns an action–effect relationship within that scaffold; it is not constructing an entire body schema from an undifferentiated sensory stream. Nor should sensorimotor contingency detection be equated with a human subjective sense of agency. The study addresses mechanisms that could support body knowledge, causal attribution, and agency. It does not show that the robot experiences authorship or selfhood. Raw prediction error remains an imperfect proxy for useful curiosity. Error can indicate an informative gap in the model, hidden environmental state, sensor noise, nonstationarity, or an intrinsically unpredictable event. Extending this approach to open environments will require distinguishing reducible uncertainty from randomness. Finally, the mobile is a deliberately small world. Real robotic agency involves simultaneous visual, tactile, vestibular, proprioceptive, social, and task-related consequences unfolding over multiple time scales. The value of the crib is not ecological completeness. It is experimental legibility. The larger lesson. A great deal of embodied AI currently begins with a task: walk here, grasp this, imitate that trajectory, maximize this reward. Robot in a crib begins one level earlier. Before useful action comes the discovery that action has consequences. The paper’s deepest contribution is to resist turning that discovery into a single number. Learning may appear as more movement, but it may also appear as better timing, a different pose, selective stillness, exploitation of environmental dynamics, or a switch from broad exploration to efficient control. That is why the crib is more than a playful visual metaphor. It is a causal microscope. By reducing the world to a body, a mobile, prediction, and curiosity, the experiment makes visible a principle that scales far beyond infancy:. An autonomous agent does not merely learn from the data the world gives it. It learns by acting, noticing what changed, and deciding what uncertainty is worth touching next.