Backpropagation is widely regarded as the canonical solution to the credit assignment problem in deep learning. However, its debated biological plausibility has motivated decades of research into local learning rules. Here we argue that whether backpropagation is required for reliable credit assignment depends less on that debate than on the structural properties of the learning problem. We formulate credit assignment as a local inverse problem and argue that when architecture restricts learning to task-relevant subspaces that are locally identifiable, reasonably well-conditioned, and supplied with appropriate local error or modulatory signals, local plasticity can suffice; when architecture fails to impose such restrictions, some form of broader coordination may become necessary. We propose that neural architecture, shaped by evolution and development, pre-structures biological learning problems so that many such subspaces are already constrained in this way. We distinguish different regimes of biological learning, including one-shot episodic storage, gradual skill refinement, and structural representational adaptation. Within this framework, associative memory supports rapid storage and context-dependent retrieval, modular organization localizes credit assignment, and replay can improve conditioning and stabilize local refinement. Backpropagation, or some comparable form of broader coordination, is therefore not universally required for learning, but may become necessary when architecture and task structure fail to confine learning to identifiable, well-conditioned subspaces.

Credit assignment in neural systems is commonly framed as the problem of computing or approximating gradients of a global objective. Under realistic constraints of partial observability, noise, delay, and high dimensionality, global credit assignment is computationally demanding and structurally ill-posed. Locality, modularity, and hierarchy can therefore be interpreted as structural responses that improve identifiability under these constraints. Biological learning does not solve the global inverse problem; it avoids it by restricting credit to locally identifiable subsystems.

Abstraction and hierarchical organization are pervasive features of biological and artificial intelligence. Such systems exhibit invariance to irrelevant variation, reuse structure across contexts, organize behavior across multiple timescales, and decompose complex actions into nested substructures. These properties are often treated as representational achievements, learned strategies, or architectural design choices. Here we argue instead that abstraction and hierarchy are coupled consequences of control feasibility in finite closed-loop embodied systems. Finite systems operate under limits of memory, processing capacity, bandwidth, time, noise, delay, and stability. When environmental complexity exceeds system capacity, systems must compress task-relevant structure in order to function. When this compression preserves relations needed for prediction, action, or regulation while discarding irrelevant variation, abstraction emerges as constrained compression. However, in closed-loop control, abstraction cannot generally remain a single-level operation. Variables that are stabilizable at one temporal or spatial scale may be unstable, unobservable, or uncontrollable at another. Hierarchy therefore emerges as the organization of compressed control-relevant variables across temporal and spatial scales, allowing fast local processes to stabilize immediate dynamics while slower higher-level processes operate on reduced variables whose dynamics have already been made tractable. On this view, hierarchy, modularity, temporal chunking, dimensionality reduction, and abstraction are not independent cognitive capacities, but related consequences of constrained processing and control. The framework offers a unifying account of why hierarchical abstraction is expected to appear across motor control, speech production, sequential action, perception, and artificial learning systems, and reframes generalization as feasibility-constrained reuse rather than representational sophistication alone.

Motor learning is shaped by history, context, and task structure. Such phenomena are commonly explained using state-space models of adaptation or contextual-inference accounts, the latter proposing that inferred latent contexts or task states select or combine internal models or control policies. Here we propose a complementary control-theoretic account grounded in control feasibility.We argue that effective motor control and learning may depend on the availability, construction, persistence, and decay of control-relevant variables: quantities available to or constructed within the controller that contribute to the generation, selection, or regulation of action through feedback, feedforward, or combined control pathways. Stabilizing variables form the subset of these control-relevant variables whose use within a controller supports stable and reliable regulation. This role is conditional, and the same variable may stabilize one controller or parameter regime yet be ineffective or destabilizing in another.Using simplified plant-coupled simulations, we examine compensatory variables and dynamic states within feedback and feedforward architectures. Crossing force-bias and velocity-linked variables with constant-load and viscous perturbations shows that benefit depends on structural matching, while retained dynamic states produce aftereffect-like behavior after perturbation removal. An augmented-state analysis identifies a finite compensatory-gain boundary beyond which the constructed state destabilizes the closed loop. A curl-field reaching model shows how delay, noise, and contextual identifiability constrain compensation.The framework provides a control-feasibility level of analysis. It suggests that effective motor control and learning may depend not only on error, reward, or contextual inference, but on whether variables are available, identifiable, timely, and compatible with stable control.

Cognitive systems are often evaluated in terms of representational or expressive capacity, yet robust learning depends on whether transformations can be identified, stabilized, and reused under delay, noise, and partial observability. Information theory characterizes what can be represented or transmitted, but not what can be stably transformed under physical constraints. We argue that standard assumptions about learnability become problematic once delay, noise, and partial observability are treated as structural constraints rather than nuisances. We develop a unified framework in which learnability depends not only on representational structure but also on structural conditions for observability and identifiability, with stability treated as a distinct closed-loop constraint. This framework is directly relevant to psychological theories of motor learning, perception, and cognitive control, in which successful behavior depends on learning transformations that remain inferable and stable under real-world conditions. We further argue that some highly expressive transformation classes can be structurally under-constrained, making task-relevant structure difficult to infer and increasing the risk of brittle solutions. Some representations may therefore be not merely difficult to learn, but effectively unlearnable from the signals available under delay, noise, partial observability, and hidden-variable coupling. These limits do not arise solely from failures of optimization, but from feasibility constraints on inference in closed-loop systems. Viewing learned transformations as constrained functional constructs helps clarify poor generalization, instability, and context specificity in motor and cognitive learning, and provides principled guidance for interpreting behavioral and computational results.

Aging is associated with changes in sensorimotor control that contribute to functional decline, mobility limitations, and increased fall risk. Traditional motor assessments often rely on subjective measures, highlighting the need for objective, quantitative tools. We developed three robot-based tasks using the vBOT planar manipulandum to evaluate sensorimotor performance in healthy young (<35 years) and older (>60 years) adults. These tasks uniquely combined bimanual control and altered dynamic conditions to assess age-related differences. The first task required bimanual coordination to control a virtual 2D arm over 400 center-out and return trials, targeting de novo motor learning. The second task involved unimanual reaching with the dominant hand, consisting of 200 trials in a null-field condition followed by 200 trials with object-like dynamic forces. The third task similarly began with 200 null-field trials and then introduced a viscous force field in the final 200 trials, with fast movements rewarded to encourage peak performance. This task also enabled comparison between dominant and non-dominant arms. All tasks detected age-related performance differences, with the viscous resistance task proving most sensitive to declines in movement speed, force generation, and response onset time. Scoring mechanisms that encouraged brisk performance amplified these effects. Across tasks, older adults generally moved more slowly, took longer to complete tasks, exerted lower peak forces, and had longer response onset times. However, some older participants performed comparably to younger individuals. In the third task, dominant arm performance consistently exceeded that of the non-dominant arm. These results demonstrate that robot-based tasks can sensitively quantify age-related sensorimotor decline and may offer valuable metrics for clinical assessment and monitoring.

Intelligent behavior across a wide range of activities such as motor control, speech production, learning, and social cognition must be generated and operate under severe uncertainty: sensory feedback is noisy and delayed, actions are imprecise, and the true causes of observable outcomes are rarely directly accessible. We argue that under these conditions, for cognition to be optimal, it cannot be organized only around direct control or evaluation of observable actions. Instead, it must largely operate through inference over latent state variables that summarize the underlying condition of the system and determine its future behavior. From this perspective, movements, sensory inputs, and social behaviors are treated as noisy evidence about hidden causes—such as motor state, articulatory configuration, or intentions—rather than as primary targets of control or judgment. We show that this state-centric organization across many domains is not a modelling preference but a functional necessity for feasible behavior under uncertainty, and that it provides a unified explanation for robustness, tolerance of variability, context sensitivity, and credit assignment across cognitive domains. Cognition, from this view, is best understood as continuous inference over latent states in noisy, embodied, and social systems.

The human motor system can learn to control novel effectors, but the contribution of task-relevant haptic dynamics to de novo learning remains unclear. Using a bimanual robotic interface, participants learned over two days to control the shoulder and elbow angles of a virtual arm in order to achieve accurate endpoint movements via constrained handle motions. On Day 1, one group practiced a purely kinematic mapping, whereas another group received continuous haptic feedback generated by an endpoint mass. With practice, movements shifted from sequential to more coordinated control and trajectories became straighter, with reduced directional deviation during target-directed endpoint movements, particularly in the haptic-feedback group. On Day 2, both groups learned to compensate for a velocity-dependent force field. Trajectories were initially curved but straightened with practice, and washout produced after-effects. Training in the presence of task-relevant haptic dynamics was associated with more complete error reduction during force-field exposure, while maintaining robust after-effects. Exponential modeling provided no evidence for a difference in learning rate between groups but was consistent with a lower residual (asymptotic) error in the haptic-feedback condition. These benefits therefore reflected a difference in final predictive compensation rather than in the speed of adaptation. Together, these results suggest that performance in the presence of task-relevant haptic dynamics was associated with more complete predictive compensation when adapting to novel dynamics, without evidence of faster adaptation.

Adaptive sensorimotor behavior relies on neural systems that reuse stable computational functions across tasks and contexts, rather than constructing control structures de novo for each behavioral demand. Biological nervous systems exhibit this organization through partially genetically specified architectures composed of reusable microcircuits that implement core computations and are refined through development and experience. With a focus on sensorimotor control, we propose a computational-level framework in which neural systems are constructed from abstract functional building blocks corresponding to reusable components such as state estimation, prediction, coordinate transformation, error mapping, memory-based control, and low-dimensional synergies. These components are task-agnostic by design yet adaptable, supporting reuse and recombination across behaviors while preserving stable computational roles. Making such functional components explicit provides a principled account of inductive bias, learning efficiency, transfer, and robustness in adaptive systems. From this perspective, persistent limitations in transfer and continual learning in contemporary artificial neural networks can be understood as consequences of treating parameters rather than computational functions as the primary units of learning. The framework therefore motivates a shift from parameter-centric optimization toward learning and recombination of reusable computational functions.

Intelligent behavior across motor control, speech production, learning, and social cognition operates under uncertainty: sensory feedback is noisy and delayed, actions are imprecise, and observable outcomes often have ambiguous causes. When multiple latent states generate identical or sufficiently similar observations, regulation, learning, and evaluation based on behavior alone become underdetermined. We argue that inference over latent state is therefore a feasibility condition for cognition under partial observability, rather than merely a modeling convenience. Observable signals—movements, sensory inputs, speech outputs, and social actions—should be understood as noisy evidence about hidden causes, including motor state, articulatory configuration, task context, and intentions. This perspective suggests that separate psychological literatures share a common control-theoretic structure: regulation under noise, delay, ambiguity, and partial observability. It also helps explain why diverse domains exhibit robustness, tolerance of variability, context sensitivity, and structured credit assignment. These properties arise from regulation of inferred, task-relevant state, rather than precise control of observable behavior. The account is not proposed as a replacement for belief-state, internal-model, predictive-coding, or active-inference approaches, but as a cross-domain feasibility argument: latent-state representations support control, learning, and evaluation only when task-relevant hidden distinctions are detectable from observations and stabilizable or actionable. Cognition, from this view, depends on latent state estimates that render behavior, learning, and evaluation well-defined under uncertainty.

Intelligent behavior across domains such as motor control, speech production, learning, and social cognition must operate under severe uncertainty: sensory feedback is noisy and delayed, actions are imprecise, and the causes of observable outcomes are often ambiguous. Under these conditions, cognition cannot be organized solely around observable behavior. When multiple latent states generate identical observations, regulation, learning, and evaluation based on behavior alone become fundamentally ill-posed. We argue that inference over latent state is therefore not a modeling choice but a necessary condition for feasible cognition. Observable signals—movements, sensory inputs, and social actions—must be treated as noisy evidence about hidden causes, such as motor state, articulatory configuration, or intentions, rather than as primary targets of control or judgment. This perspective explains why diverse domains exhibit robustness, tolerance of variability, context sensitivity, and structured credit assignment. These properties arise not from precise control of behavior, but from regulation with respect to inferred, task-relevant state under conditions of partial observability. Cognition, from this view, is organized around maintaining and updating latent state estimates that render behavior, learning, and evaluation well-defined under uncertainty.

Reinforcement learning frameworks typically formalize reward as a scalar evaluative signal that defines task objectives and drives behavioral adaptation. In simple control problems, reward can often be specified as the negative of a designer-specified, often heuristic, error or cost, making reward maximization closely related to error minimization. In complex embodied action domains, however, task-relevant evaluation is often difficult to define or measure directly. Partial observability, delayed consequences, physical interaction, and the risk of catastrophic failure mean that the variables over which success, error, or reward are computed must first be inferred, stabilized, and represented in a form that can guide control. The central proposal is that reward should be understood not merely as a predefined scalar learning signal, but as a post-inferential representation of evaluation. Prediction error, information gain, intrinsic reward, and the interpretation of extrinsic reward all presuppose a representational space within which deviation, improvement, or task success can be computed. In embodied systems, this space is often not given in advance in task-aligned form, but must be constructed through state estimation, prediction, outcome interpretation, and control. Usable reward therefore depends on prior inference and stabilization of control-relevant state variables. This reframing clarifies both the strengths and limitations of reward-based learning. It suggests that a central problem in complex action domains is not simply optimization over reward signals, however many are combined, but the construction of control-relevant state representations relative to which evaluation and credit assignment become possible.

Most studies of motor learning focus on adaptation, which can be described as recalibration within an existing controller following changes in dynamics, kinematics, or sensory feedback. In these paradigms, control variables, coordinate representations, and feedback organization are treated as available a priori, and learning as parameter tuning within a specified controller architecture. However, de novo motor-learning tasks require learners to establish novel relationships between intention, sensory feedback, and action, rather than merely recalibrating a familiar control organization. These tasks often show slow or variable acquisition, selective and task-dependent generalization, strong context dependence, and dissociations between learning rate and final performance. Such features are difficult to explain solely as parameter tuning within a fixed controller architecture. Here we propose that de novo motor learning is better understood as controller synthesis, rather than only as parameter adaptation. On this view, learning involves forming and stabilizing content-addressable, plant-state-addressed controller memories: local sensorimotor control organizations that specify where a controller applies, which task-relevant signals are selected and routed, how state and progress are estimated, how control is computed, how internal control signals map onto action, and when and how the controller is expressed. This framework interprets slow or variable acquisition, structured generalization, context dependence, and learning-rate/outcome dissociations as possible consequences of controller-memory formation, stabilization, retrieval, and expression, rather than of slow parameter convergence within an established controller. It also suggests tests for distinguishing controller synthesis from fixed-structure alternatives based on adaptation, contextual inference, or policy learning.

The human motor system exhibits remarkable plasticity: not only can we master complex skills, but we can also learn to control artificial effectors. Here, we examine whether haptic force feedback from a simulated endpoint mass facilitates the de novo learning of novel kinematic and dynamic mappings. We investigate this using a virtual 2D planar arm controlled via a bimanual robotic manipulandum. Human participants moved two handles constrained to 1D channels, with handle positions specifying shoulder and elbow joint angles. Across two days, they performed center-out and out-back movements under two experimental conditions. On Day 1, one half of the participants practiced with a purely kinematic arm, whereas the other half received continuous haptic feedback from an endpoint mass. Although intrinsic hand movements were often initially sequential, participants rapidly adopted more coordinated control strategies, and trajectories became straighter with practice. Notably, the haptic-feedback group exhibited significantly greater improvements in virtual arm control. On Day 2, we introduced a velocity-dependent curl field to both groups. The curl field initially resulted in curved, loopy paths, but participants gradually compensated, consistent with internal-model formation. Prior exposure to haptic feedback conferred a superior capacity to adapt to the novel dynamics, evidenced by straighter trajectories and smaller directional biases throughout force-field exposure. These findings suggest haptic feedback enhances sensorimotor learning, possibly by engaging a distributed neural network of brain regions involved in motor skill acquisition and refinement. Significance Statement Humans can learn to control arbitrary extensions of the body. However, the factors that influence such de novo motor learning remain unclear. Here, we demonstrate that haptic force feedback from a simulated endpoint mass enhances the acquisition of novel kinematic and dynamic mappings using a virtual arm. Participants who trained with haptic feedback showed greater adaptation to force fields and more accurate trajectory control. These results suggest that haptic input facilitates the formation of more efficient internal models during novel sensorimotor learning. In addition to their theoretical implications, these findings have practical implications for the use of haptic feedback during training for rehabilitation and the control of complex machinery. ### Competing Interest Statement The authors have declared no competing interest.

Contextual cues arising from distinct movements are crucial in shaping control strategies for human movement. Here, we examine the impact of visual and passive lead-in movement cues on unimanual motor learning, focusing on the influence of "dwell time," where two-part movements are separated by the interval between the end of the first movement and the start of the second. We used a robotic manipulandum to implement a point-to-point interference task with switching opposing viscous curl fields in male and female human participants. Consistent with prior research, in both visual and passive lead-in conditions, participants showed significant adaptation to opposing dynamics with short dwell times. As dwell time increased for both visual and passive signals, past movement information had less contextual influence. However, the efficacy of visual movement cues declined more rapidly as dwell times increased. At dwell times greater than 800 ms, the contextual influence of prior visual movement was small, whereas the effectiveness of passive lead-in movement was found to be significantly greater. This indicates that the effectiveness of sensory movement cues in motor learning is modality dependent. We hypothesize that such differences may arise because proprioceptive signals directly relate to arm movements, whereas visual inputs exhibit longer latency and, in addition, can relate to many aspects of movement in the environment and not just to our own arm movements. Therefore, the motor system may not always find visual movement cues as relevant for predictive control of dynamics.NEW & NOTEWORTHY This research uncovers, for the first time, how visual and proprioceptive sensory cues affect motor learning as a function of the pause or "dwell time" in two-part movements. The study has shown that visual lead-in movement cues lose their effectiveness sooner than passive lead-in movement cues as dwell time increases. By revealing the modality-dependent nature of sensory information, this study enhances our understanding of motor control and opens new possibilities for improving therapeutic interventions.

Motor adaptation is typically studied using simplified virtual tasks. Here, we investigated how humans learn to stabilize a physically unstable, underactuated system, and how participants cope with changes in the system’s dynamics. Twelve right-handed adults balanced a real inverted pendulum by moving a cart along a linear rail. Study 1 characterized the passive mechanical properties of three pendulums (short, medium, and long) using free-oscillatory decay and fall to ±30° trials, after their release from the upright position. Longer rods exhibited slower decay and lower natural frequencies, as well as a longer duration before falling, indicating greater passive stability. Study 2 assessed human motor control of the pendulum to maintain balance. Human participants trained with the medium pendulum (30 trials) and were then tested with all three pendulums (20 trials each). During training, balance performance improved significantly, with time to failure increasing over trials. During testing, performance scaled with pendulum length, and longer rods were easier to balance. Similar peak cart velocities were observed across conditions, suggesting equivalent actuation effort. Additionally, as expected, pendulum angular velocities decreased with rod length, reflecting underlying inertial differences. Pendulum passive dynamics closely matched behavioral performance, supporting a strong link between intrinsic system properties and balancing outcomes. These findings show that motor learning in physically unstable environments is not only shaped by feedback and effort, but also by the alignment of human control strategies and abilities with the natural dynamics of the plant. We note that in this study, we used a modified pendulum rig previously employed to examine control engineering approaches to modelling balance, thereby generating a dataset that can later be used to compare human performance with real-time computer control implementations of the same tasks. ### Competing Interest Statement The authors have declared no competing interest.

Motor imagery is frequently utilized to improve the performance of specific target movements in sports and rehabilitation. In this study, we show that motor imagery can facilitate learning of not only the imagined target movements but also sequentially linked overt movements. Hybrid sequences comprising imagined and physically executed segments allowed participants to learn specific movement characteristics of the executed segments when they were consistently associated with specific imagined segments. Electrophysiological recordings revealed that the degree of event-related synchronization in the alpha and beta bands during a basic motor imagery task was correlated with imagery-evoked motor learning. Thus, both behavioral and neural evidence indicate that motor imagery's benefits extend beyond the imagined movements, improving performance in linked overt movements. This provides decisive evidence for the functional equivalence of imagined and overt movements and suggests applications for imagery in sports and rehabilitation.

We describe the construction and evaluation of two robotic grippers for berry picking. Using a pneumatic cylinder drive, one was constructed from hard materials and the other from soft materials. A novel evaluation paradigm using a handle mechanism was developed, so the grippers could be directly op-erated by human participants. An artificial bush was also constructed and used for evaluation purposes. Overall, both grippers performed worse than the human hand, indicating that further development is needed.

In daily life, we coordinate both simultaneous and sequential bimanual movements to manipulate objects. Our ability to rapidly account for different object dynamics suggests there are neural mechanisms to quickly deal with them. Here we investigate how actions of one arm can serve as a contextual cue for the other arm and facilitate adaptation. Specifically, we examine the temporal characteristics that underlie motor memory formation and recall, by testing the contextual effects of prior, simultaneous, and post contralateral arm movements in both male and female human participants. To do so, we measure their temporal generalization in three bimanual interference tasks. Importantly, the timing context of the learned action plays a pivotal role in the temporal generalization. While motor memories trained with post adaptation contextual movements generalize broadly, motor memories trained with prior contextual movements exhibit limited generalization, and motor memories trained with simultaneous contextual movements do not generalize to prior or post contextual timings. This highlights temporal tuning in sensorimotor plasticity: different training conditions yield substantially different temporal generalization characteristics. Since these generalizations extend far beyond any variability in training times, we suggest that the observed differences may stem from inherent differences in the use of prior, current, and post adaptation contextual information in the generation of natural behavior. This would imply differences in the underlying neural circuitry involved in learning and executing the corresponding coordinated bimanual movements.