A person lifts a cup that is unexpectedly empty. The hand rises too quickly, then corrects. Before the lift, the nervous system had prepared for a different weight and anticipated a different sensory consequence. Sensorimotor prediction is the process of using a motor plan and past experience to estimate what a movement will feel like and what it will do. It makes rapid action possible despite delayed sensory feedback. The prediction is useful because it can be wrong: the resulting error helps the system revise the next action.
What is predicted during movement?
A motor command changes the body and the world. The relevant sensory consequences can include joint position, muscle force, vision, touch, and timing. A forward model estimates a future state or expected feedback from the current state and planned command. Wolpert, Ghahramani, and Jordan’s internal-model work formalized this approach to sensorimotor integration. The model is a computational explanation, not a claim that there is one literal miniature simulator in the brain. Different movements, sensory channels, and contexts may rely on partly distinct predictive processes.
Prediction matters because feedback arrives after the command has begun. If a person waits for the hand to miss the cup before deciding how to move, many fast actions become unstable. Anticipation can guide force and trajectory in advance, while incoming sensory evidence adjusts the estimate. The cup example combines these functions: an expected weight shapes the initial lift; the actual movement provides evidence that the estimate was wrong. The next lift may use a different force. This does not mean prediction replaces feedback. Effective control uses both.
Sensorimotor prediction is narrower than the idea that the brain predicts everything. The article concerns consequences of action and state estimation in movement. A reward prediction error concerns expected value; a sensory prediction in perception can occur without a motor command. These concepts may interact, but their experiments and claims differ. A useful explanation says what was predicted, which signal differed, and which behavior changed afterward. Otherwise the word prediction becomes too broad to test.
Forward models and sensory error
Shadmehr, Smith, and Krakauer reviewed how forward models can predict sensory consequences and support adaptation. When actual feedback differs from expected feedback, the mismatch can help update motor commands. In reaching experiments, a shifted visual display or altered force field makes this visible: movements initially miss, then change with practice. Such adaptation is evidence of learning under a particular perturbation. It is not proof that one error-correction rule explains every motor skill, nor that every improvement is driven by a conscious sense of surprise.
A prediction error is not always a simple difference between two measurements. The system needs an estimate of body state, and each sensory source has uncertainty. Vision may be precise about location but delayed; proprioception may be faster for some properties yet less precise in another context. Körding and Wolpert’s sensorimotor-learning experiments examined behavior consistent with combining prior information and uncertain sensory evidence. A Bayesian account describes how weighting could depend on reliability. It is a model to test, not a claim that participants consciously calculate probabilities while moving.
Noise makes prediction especially valuable. Muscles do not produce an identical outcome every time, objects change, and the environment can move. A robust controller estimates likely consequences and corrects when evidence arrives. An error can reflect an inaccurate internal estimate, a changed environment, or random variation. A study of learning must distinguish systematic adaptation from momentary correction. If the cup was empty only once, the next action might be cautious without a durable new model. Repeated exposures can reveal whether the change persists.
Predictions can be about timing as well as force or position. A hand opening to catch a ball must be prepared before contact, and the predicted contact time can shape grip. If the ball arrives early, corrections may be possible only after a delay. Tasks that vary arrival time can reveal how tightly anticipation is coupled to action. They also show why a late correction is not necessarily a weak predictive system; the available feedback and biomechanical constraints determine what could have been changed after the surprise.
The same movement can be controlled at several levels. A person can decide to reach for a cup, choose a route around an obstacle, and regulate fingertip force as the cup is lifted. A forward model for contact dynamics would not replace the goal-level decision, and a goal-level expectation would not specify every muscle command. Describing the level of prediction prevents a motor-control account from swallowing all planning. It also helps explain why learning at one level may transfer differently from learning at another.
Why self-generated sensations can feel different
One proposed function of sensorimotor prediction is to distinguish consequences of one’s own actions from unexpected external events. Blakemore, Wolpert, and Frith studied why self-produced tickle is typically less intense than similar externally produced touch. Their experiment supports a relation between predicted self-generated sensation and attenuation under the tested conditions. It does not show that every self-produced sensation is cancelled, or that an attenuated sensation always reveals a specific brain computation. Timing and the match between action and feedback are crucial.
The popular explanation “the brain cancels what it predicts” is too sweeping. Self-generated touch remains detectable and sometimes highly salient. The useful claim is that predictions can alter sensory processing when the actual input matches expected consequences. An unpredictable delay or spatial mismatch can change the experience. Experiments can manipulate timing and agency to test the account, but participants may also change attention and expectations. Inference requires controls for those alternatives.
Self–other distinction is related but not identical. Knowing that a movement was one’s own involves intentions, action monitoring, and context in addition to sensory prediction. A person may predict a visual consequence of a robot’s movement without having authored the movement. Conversely, a self-generated action may have a surprising consequence. A mechanism guide should resist translating a specific sensory attenuation result into a complete theory of agency or responsibility.
Neural pathways and the cerebellum
The cerebellum is a prominent candidate in forward-model accounts because of its connectivity and role in coordinated movement and adaptation. Tanaka and colleagues reviewed evidence for cerebro-cerebellar predictive computation while noting that there is no complete consensus about its specific function. A region associated with adaptation may contribute to timing, error processing, or other operations as well. Lesions can support causal inference but can disrupt several pathways. The strongest claim is about converging patterns from behavior, physiology, and carefully selected neural studies.
Other circuits also matter. Motor planning involves cortical and subcortical networks, and sensory state estimation uses incoming signals distributed across several systems. A single anatomical label cannot explain how the hand anticipates a cup’s weight. Computational models help specify the information that must be available, while neural experiments test possible implementations. The two levels should inform each other without being confused. A model can predict behavior well and still be implemented differently from its diagrams.
Neural evidence is particularly easy to overstate when a task changes several things at once. If a participant learns a new force field, they also attend, adjust speed, and experience fatigue. An imaging difference after learning may reflect prediction updating, altered movement, or changed effort. Control tasks and behavioral readouts narrow the interpretation. Repeated experiments across modalities and perturbations are needed before a mechanism is generalized beyond the original reaching setup.
Measurement and task design
Researchers commonly infer prediction from patterns of movement, force, sensory attenuation, and aftereffects. An aftereffect can be informative: when a perturbation is removed, a learned compensation may briefly persist, showing that action preparation changed rather than only reacting online. Yet aftereffects depend on the task and do not expose every internal variable. A participant can consciously adopt a strategy or unconsciously adapt, with different implications for transfer. The companion measurement article considers these construct-validity problems in detail.
Timing matters at millisecond scales. If a hand corrects before a slow visual signal could plausibly have guided it, prediction is a candidate explanation. But other sensory channels may provide earlier information. Experiments need realistic estimates of delay and should record trajectories, not only final success. Two participants may place a cup in the right position while using different correction strategies. Endpoint accuracy alone cannot show whether a forward model was used.
Movement variability is another clue. If a person learns that an object is uncertain in weight, they may grip more firmly or move more slowly. That adaptation could reflect rational management of uncertainty rather than a failure to predict. A study should report speed, force, and error together. It should also consider fatigue and practice effects, which can change all three. A single “prediction score” hides these interactions unless it is tied to a precise computational model and validated tasks.
Researchers can strengthen inference by comparing conditions that differ in predictability but are matched for basic movement demands. If the hand follows the same path while the feedback becomes delayed or shifted, changes in adaptation are more interpretable than if the task changes in many ways at once. Even a matched design needs checks: participants may notice the manipulation and adopt an explicit strategy. Asking about awareness, recording trial-by-trial adjustment, and testing aftereffects can help distinguish processes without pretending that a single behavioral signature is perfectly pure.
Applications without a promise of universal transfer
Rehabilitation, sports training, and interface design all have an interest in how people anticipate action outcomes. A haptic interface that responds late may feel awkward because the feedback no longer matches the user’s movement. A tool with consistent weight and response may become easier to handle. These examples are Institute synthesis from sensorimotor principles; they are not claims that the cited laboratory studies validated a particular device or therapy. Any application must be tested with the intended users and their specific tasks.
Practice can improve a trained movement, but generalization is an empirical question. Adaptation to one visual shift may not make someone broadly better at all hand-eye coordination. The nervous system can learn task-specific mappings. A course or product should avoid promising that generic prediction drills improve every movement or decision. Transfer requires a reason to expect shared structure and a test outside the practiced task. The same standard applies to claims about cognition built on motor-learning metaphors.
Prediction can also protect against surprise while leaving room for exploration. A rigid expectation that never updates would make a person clumsy when objects change. Accurate control balances prior experience with new sensory evidence. This is an appealing design principle, but its exact weighting depends on reliability and cost in the particular environment. A fragile object may justify caution even if it is usually light; a high-speed movement may rely more strongly on anticipation because feedback arrives too late. The optimal policy is contextual rather than a fixed amount of prediction.
For people adapting to a changed body or assistive device, the context can change substantially. A prosthetic hand may provide different sensory feedback from a biological hand, so familiar predictions may be less reliable at first. Training and interface design can support learning, but the cited basic studies do not tell us the best rehabilitation protocol. Evaluation should ask whether a person improves on meaningful everyday tasks, how comfortable the device is, and whether gains remain outside the laboratory. A compelling forward-model explanation is a starting hypothesis, not an outcome measure.
That distinction also matters for measurement over time. A better score on the device used during training may reflect familiarity with that exact device. Testing a different task, observing use at home, and asking the user which actions became easier would reveal more about practical transfer. Improvements in predictive control may be one part of the story, alongside changing confidence, effort, and environmental support.
What the evidence does not show
The evidence does not show that one forward model predicts every bodily sensation, that a cerebellar signal proves a conscious expectation, or that all changes after practice reflect the same prediction-error mechanism. It does not establish a universal training method for movement or cognition. Experimental work supports predictive state estimation and adaptation in defined motor tasks, while leaving questions about implementation, transfer, and individual variation. Claims should name the action, signal, perturbation, and measured response.
Sensorimotor prediction explains how action can be prepared for a world that is not fully known in advance. It joins an expectation from prior experience with feedback from the movement actually made. When the cup is lighter than expected, the error is useful information, not simply a failure. The scientific value of the idea lies in making that cycle testable and in resisting the urge to turn one elegant model into an explanation for every human capacity.
Research references
- Wolpert, D. M., Ghahramani, Z., & Jordan, M. I. (1995). An internal model for sensorimotor integration. Science, 269(5232), 1880–1882. View source
- Shadmehr, R., Smith, M. A., & Krakauer, J. W. (2010). Error correction, sensory prediction, and adaptation in motor control. Annual Review of Neuroscience, 33, 89–108. View source
- Blakemore, S.-J., Wolpert, D. M., & Frith, C. D. (1998). Central cancellation of self-produced tickle sensation. Nature Neuroscience, 1, 635–640. View source
- Körding, K. P., & Wolpert, D. M. (2004). Bayesian integration in sensorimotor learning. Nature, 427, 244–247. View source
- Tanaka, H., Ishikawa, T., Lee, J., & Kakei, S. (2020). The cerebro-cerebellum as a locus of forward model: A review. Frontiers in Systems Neuroscience, 14, 19. View source