The Nervous System · Part Two · How It Senses and Moves
Lesson 36 / 61
Internal Models and the Motor Hierarchy: How the Brain Steers a Limb Before Feedback Arrives
The signal that says where your arm is arrives after the arm has already moved.
Motor control is the circuitry that turns an intention into muscle contraction and keeps the result accurate. Every signal reporting where a limb is arrives late, so the nervous system steers on a forward model that estimates the limb's state before the body can confirm it. The Unified Model of Tone identifies a mechanical input as information delivered to that estimate, which is why the segment it reaches decides what changes.
Feedback lag
Receptor transduction and conduction make even the fastest cortical loop too slow to steer a reach
Adaptation target
Practice inside a force field rebuilds the model of the limb's dynamics and leaves the kinematic plan intact
Module weighting
Each pair's contribution scales with how accurately its own forward model predicts
Learning modes
Supervised in cerebellum, reinforcement in basal ganglia, unsupervised in cortex
Motor control
Primary motor cortex, premotor cortex and the supplementary motor area issue the command. The corticospinal, reticulospinal and vestibulospinal tracts carry it down to spinal motoneurons. The cerebellum and basal ganglia shape the command before it leaves and correct the movement while it runs. Fine finger control travels corticospinally, and posture and gait travel by the brainstem routes.
Movement and tone
Tone is the organization the nervous system holds across everything it runs, and in movement that organization shows up as the accuracy of the body's report of itself. The forecast a limb is steered by is only as good as that report. Vibrate the neck muscles of a blindfolded walker and the trajectory bends away from the target. Nothing was done to the legs.
01Prediction in motor control
Motor control begins with a forecast because feedback arrives too late to steer
The nervous system has to know the position and velocity of every segment it controls, and every signal that reports them arrives late, delayed by receptor transduction and conduction along the nerve. After an imposed displacement of the wrist in ten healthy adults, the area 4 dipole moment began at 35 ms and peaked at 54 ms MacKinnon 2000. That timing suits synaptic activity onto corticospinal neurons driving the long-latency response, and it is far too late to steer a fast reach.
The answer is a forward model, an internal estimator that predicts the sensory consequences of a motor command before the body can confirm them. Miall proposed that the cerebellum builds two of them. One is a forward predictive model of the limb and muscle. The second is a model of the delays in the loop itself, arising from receptor and effector lag, axonal conduction and cognitive processing Miall 1993. Modeling the delay is what lets a prediction and a late-arriving sensation be compared in the same time frame.
Behavioral work corroborates the estimator. Error in people's judgments of an unseen hand propagates over time in the shape an optimal state estimator predicts Wolpert 1995, a demonstration worked through on the research page on predictive processing.
Running ahead of feedback defeats the delay and carries a cost. An imperfect model drifts, accumulating error with every cycle it is left uncorrected. Prediction error is the training signal that pulls it back, which is why the system keeps comparing the forecast against what actually arrives.
Practice rebuilds the dynamics and leaves the plan alone
Human subjects reached while a robot manipulandum imposed a velocity-dependent force field, and their first movements were grossly distorted. With practice the hand paths inside the field converged on trajectories very close to the ones drawn in free space Shadmehr 1994. The kinematic plan survived the change in dynamics. What the nervous system rebuilt was the model of the forces, and not the shape of the reach.
Removing the field exposed what had been built. Trajectories became approximate mirror images of the first distorted attempts, and those aftereffects reached workspace regions where no exposure to the field had taken place Shadmehr 1994. Transfer beyond the training data rules out a lookup table of visited states and experienced forces. The motor system had learned the dynamics, and it held them in a coordinate system close to that of the joints and muscles.
02Findings
What the research shows
03Efference copy and self
Efference copy lets the brain subtract the sensation its own movement predicts
Every motor command leaves a copy of itself behind, and that copy predicts the sensation the movement is about to cause. The nervous system subtracts the predicted reafference from the incoming stream, canceling the sensory effects of self-motion so genuinely external events stand out. It is why a stable world does not smear as the eyes sweep across it.
Imaging localized the subtraction. Identical tactile stimulation produced more somatosensory cortex activity when it was externally produced than when subjects produced it themselves. Cerebellar activity was lower for a movement that generated a predictable tactile consequence than for a movement that generated none Blakemore 1998. The cerebellum carries the signal that marks a sensation as one the movement already accounted for. How far that attenuation is graded by delay and by spatial mismatch, and how it is measured in patients, is set out on the research page on predictive processing.
Cancellation answers a second question while it works, which is who moved. Comparing predicted reafference against delivered reafference is what assigns a motion of the body to the self or to an external agent. A single touch reads that comparison at the scale of one contact. Ownership of a gesture reads the same comparison at the scale of a whole limb, and that is where parietal cortex enters.
Parietal damage separates two kinds of prediction
Left parietal damage blurs the boundary between self and other. Three apraxic patients with left parietal lesions executed finger movements with either hand while the visual feedback they received was manipulated systematically. They could not reliably reject the examiner's hand as their own, and even when they moved inaccurately they claimed the correct movement on the screen was theirs Sirigu 1999. The failure appeared for complex gestures and spared simple ones. The authors read parietal cortex as maintaining a kinesthetic model of the movement in progress.
A later study sharpened what parietal cortex holds. Patients with parietal lesions could report when they started moving. They could not report when they first became aware of the intention to move. Cerebellar patients behaved like healthy subjects on the same measure Sirigu 2004. Prediction for awareness and prediction for correction are separable, and they sit at different addresses.
04Modular internal models
The motor system runs paired predictors and controllers and weights them by accuracy
The brain does not hold one controller. Under the Modular Selection and Identification for Control framework, the MOSAIC model, the motor system runs multiple pairs of models at once. Each pair holds an inverse model that generates a command and a forward model that predicts what that command will produce. The two are acquired together through motor learning, and the forward model's predictive accuracy decides how much its partner controller contributes to the final motor command Wolpert 1998.
Selection is competitive and continuous. As movement unfolds the brain compares each module's predicted consequences against actual sensory feedback. Modules with the least error are weighted most heavily, so the system identifies which internal model best matches the present world and acts through it.
The MOSAIC architecture learns objects and generalizes across them
Computer simulations of an object manipulation task showed the architecture learning to handle several objects and switching between them appropriately. It generalized to novel objects whose dynamics lay inside the range of those already learned Haruno 2001. This is how a single arm grips a feather and an anvil without conscious recalculation. Context estimation runs beneath awareness and swaps controllers as the task demands.
The failure mode is as informative as the success. When the same simulation met a novel shape paired with unexpected dynamics, inappropriate modules activated first and online correction followed Haruno 2001. The motor system mispredicts in a specific and recoverable way, then repairs the misprediction out of the error it has just made.
New modules are acquired and then stored
Human imaging caught both stages while subjects learned a new tool, a computer mouse carrying a novel rotational transformation. Activity spread over wide areas of the cerebellum was precisely proportional to the error signal guiding acquisition. A second signal confined near the posterior superior fissure remained after learning, once error levels had been equalized across conditions, and it tracked the acquired internal model of the tool itself Imamizu 2000. Each module can be refined by the errors it makes, and new modules can be learned for objects the hand has never held.
05Motor cortex and descending tracts
Corticospinal, reticulospinal and premotor systems each carry a different part of the movement
The descending systems divide by what they control, and that division survives being tested at the level of single cells. Corticospinal output from primary motor cortex governs fine finger control. Reticulospinal and vestibulospinal output anchors gait and balance.
Rabies-virus tracing from single muscles in monkeys found two subdivisions of primary motor cortex. A rostral region lacks cortico-motoneuronal cells and matches the standard motor cortex of many mammals. A caudal region holds shoulder, elbow and finger cortico-motoneuronal cells with monosynaptic access to spinal motoneurons Rathelot 2009. Those cells bypass spinal integration and sculpt the command at the motoneuron itself.
The same division shows in human behavior. A startling sound releases a prepared movement early. That StartReact effect was significantly larger for bilateral shoulder abduction than for bimanual finger abduction, with sternocleidomastoid activation more frequent in the proximal task Maslovat 2023. Reticulospinal drive scales up for proximal, postural and bilateral acts and down for individuated fingers.
Premotor cortex splits by what it is preparing
Of 211 dorsal premotor neurons recorded in monkeys preparing to reach, 53 percent were tuned to both target location and which arm would be used. Of 109 ventral premotor neurons, 60 percent coded target location only and showed no arm selectivity Hoshi 2002. Dorsal premotor cortex prepares the action, including which limb performs it. Ventral premotor cortex prepares the target.
Ventral premotor cortex controls hand movements directed at real objects within reach. Dorsal premotor cortex drives movements triggered by arbitrary environmental cues such as visual shapes or auditory tones. The supplementary motor area performs learned sequences in the absence of any external cue, running the act from memory. In three monkeys performing sequential reaches, more than half of SMA neurons were preferentially or exclusively active for internally guided sequences, 55 percent before movement and 65 percent during it. Premotor neurons favored the visually guided sequences, 55 percent before movement and 64 percent during it Mushiake 1991. Sequence-specific neurons concentrated in SMA and transition-specific neurons in premotor cortex. Those authors warned that a strict functional dichotomy is not acceptable.
Inactivation sharpens the warning. In two monkeys trained on both tasks, comparable numbers of dorsal premotor neurons were active during visually guided and memory-guided sequences. Inactivating the area had a marked effect only on the sequences guided by memory Ohbayashi 2016. Activity in a region is one measurement and necessity is another. The cue-triggered label for dorsal premotor cortex holds for what its cells do and fails for what its loss reveals.
Primary motor cortex intermingles the representations of neighboring body parts
Within-limb somatotopy in primary motor cortex is neither spatially discrete nor sequentially ordered. Schieber gathered six features that constrain the map Schieber 2001. Output converges and diverges. Horizontal interconnections spread activity across it. Activation is distributed, lesions produce characteristic deficits, and the cortex can reorganize. Representations of any two smaller body parts overlap extensively beneath gradual gradients. Outputs to flexor, extensor and synergist muscles spread across one another. The classic homunculus data were themselves consistent with distributed organization, so the drawing was over-read. The frontal lobe holds these fields inside its larger architecture.
06Three learning modes
Three forms of learning tune the motor system, each with its own structure
Supervised learning sits in the cerebellum, reinforcement learning in the basal ganglia, and unsupervised learning in the cerebral cortex, a division Doya drew across the three structures Doya 2000.
Unsupervised learning strengthens the association between stimuli and responses through simultaneous neuronal firing, Hebbian plasticity, and runs mainly in cortex. Supervised learning improves performance by minimizing prediction error signals and runs mainly in the cerebellum, the structure best placed to correct the forward model's drift. That is why cerebellar activity during tool learning tracks the error signal so closely Imamizu 2000. Reinforcement learning raises the likelihood of a behavior after reward and lowers it after punishment, the trial-and-error tuning that the basal ganglia and dopamine broker.
None of the three structures is motor-only. The same learning logic runs in cognitive domains, which is why cerebellar and basal ganglia signatures turn up in tasks with no limb in them Doya 2000. The order in which these systems come online in a child is laid out in the pediatrics material on movement.
Cerebellar disease degrades the forecast and the system pays for it
Seven adults with cerebellar disease and seven controls reached to targets, and the patients could not produce muscle torques that predicted, accommodated and compensated the interaction torques one moving joint imposes on the next. The authors named that inability an important cause of the classic kinematic deficits of cerebellar reaching Bastian 1996. Curved paths, decomposition and overshoot are a forecasting failure. Eight further subjects with unilateral or bilateral cerebellar pathology made cyclic vertical arm movements holding an instrumented object, alongside eight healthy controls. Grip force still scaled to load with no time lag, so feedforward control survived. The fine coupling of grip force to inertial load fluctuation was significantly weakened, and the patients gripped harder throughout Rost 2005. A system holding a degraded forecast pays for it by raising its safety margin. Beneath the cortex the cerebellum compares the sensory input it receives against the outputs that caused it, while the basal ganglia automatize movement by reading cortical input and predicting likely future actions.
The vocabulary of disordered movement is built on the same systems
The words clinicians use for disordered movement are built out of these same predictive systems. Phenomenology names the observed presentation, the empirical phenotype of a movement. Diagnosis names the syndrome assembled from all those phenotypic features.
The prefixes encode the deviation. Brady means slow. Hyper and hypo mean increased and decreased. Dys means abnormal, and a means without. They fuse to suffixes such as kinesia for movement and tonia for muscle tone. Read together, that vocabulary lets a clinician describe disordered movement precisely, and it sits directly on top of the predictive systems that generate normal motion.
07Neck input and state estimation
The state estimate the motor system steers by is built partly from the neck
The upper neck is one of the densest sources of the signal a forward model is corrected against. Quantitative counts in human fetal specimens found 190 muscle spindles per gram in obliquus capitis superior, 242 per gram in obliquus capitis inferior, and 98 per gram in rectus capitis posterior. No tendon organs were seen Kulkarni 2001. Those suboccipital muscles are built as sense organs reporting length and position, not force, which is why proprioception from the upper neck weighs so heavily in the estimate.
Corrupt that report and whole-body behavior bends. Blindfolded subjects walking to a target remembered at about 4 meters undershot it and deviated toward the side opposite the vibrated neck muscles. Vibration applied before the walk left trajectory length intact and instead slowed velocity and produced non-systematic deviation Bove 2001. The plan and the execution bend in different ways. In the Unified Model of Tone, cervical proprioception is part of the internal reference the forward model steers by.
The segment chosen decides whether a spinal input reaches the cortex
In adults with recurrent neck pain, a cervical thrust delivered to a segment the clinician judged non-relevant left the N30 sensorimotor integration potential where it was, at p equal to 0.757 Niazi 2024. The force was the same in both arms of that trial. The evoked-potential evidence is taken apart in the lesson on the frontal lobe, and the claim that the analysis is the variable belongs to the research page on tone.
This is the machinery the Unified Model of Tone builds on. The forecast the motor system runs on is assembled from what the body reports about itself, and a segment holding a distorted mechanical state feeds a distorted report into that assembly. In this model a correction is information delivered to a state estimate, and not force applied to a joint. Force sets how much mechanical information arrives. Address sets whether the state estimate can use it, which is the whole of the specificity claim and the reason the two arms of that trial differed in site and not in load.
Chiropractic care works on the cervical receptors whose signal the state estimate is assembled from. The neck vibration work sets the dose. Raising the vibration frequency raised the amplitude of the deviation from the planned path Bove 2001, so distortion in the report scales the error in the walk. A forecast corrected against a truthful report keeps the limb free to arrive where it was aimed, and that freedom is the health. A motor system running on a distorted report is not weak. It spends full accuracy on a body that is not there, and that is what illness looks like from inside the motor hierarchy.
The command leaves with a prediction attached, and what comes back is only what the prediction failed to cover.
08Tone
How this system expresses tone
Motor control expresses all of tone. Here it is measured in the forecast the limb is steered by, in the mechanical coupling between segments, and in the fidelity of the report that forecast is corrected against.
Prediction
The motor system acts on a forecast of the limb. Cortical potentials after an imposed wrist displacement begin at 35 ms, far too late to steer a fast reach.
Coupling
Moving one joint torques the next. Seven adults with cerebellar disease could not forecast those interaction torques, and their reaches curved, decomposed and overshot.
Input quality
Human fetal suboccipital muscles carry up to 242 spindles per gram and no tendon organs. Vibrate a blindfolded walker's neck and the trajectory bends away from the target.
Each foundation is read against one test, whether the limb arrived where the forecast said it would. Gain: eight patients with cerebellar pathology kept feedforward scaling and gripped harder throughout, a system turning its safety margin up when the forecast degrades. Oscillation: locomotion is a rhythm the forecast steers, and neck vibration during a walk bends the trajectory while the same vibration beforehand only slows the velocity. Time course: cerebellar error activity fades as a new tool is learned while the signal carrying the acquired model persists after error has been equalized. Load: a misassigned module costs an online correction, work the system would never have spent had the forecast fit the object. Set point: after training in a force field the arm produces mirror-image aftereffects, still holding the dynamics it learned and defending them. Constraint: only the caudal subdivision of primary motor cortex holds cortico-motoneuronal cells, so monosynaptic finger control reaches exactly as far as that population does.
09Across the library
How this page relates to the rest of the library
What moving does to the body's own information about itself, and why thousands of easy repetitions leave a motor map exactly where they found it.
The unseen arm in the dark, the graded attenuation of self-produced touch, and what a standing mismatch between prediction and arrival costs in energy.
Why the segment a clinician selects is the variable under test, and what analysis changes about the input that follows it.
The structure that carries the error signal and then stores the finished internal model, taken apart layer by layer and nucleus by nucleus.
The subcortical gate that releases or restrains a movement, and the reinforcement learning that decides which program is permitted to run.
How the predicting motor system is assembled in the first year, from head control through reaching to independent gait.
The same forward model under time pressure, where the forecast has to be right before the ball arrives rather than after.
10Frequently asked
Questions about this topic
What is a forward model in motor control?
A forward model is an internal estimator that predicts the sensory consequences of a motor command before the body can confirm them. It exists because feedback is late. After an imposed wrist displacement in ten healthy adults, the earliest motor cortex response began at 35 ms and peaked at 54 ms, too slow to steer a fast reach. The cerebellum is proposed to build two such models, one of the limb and one of the loop's own delays, so prediction and sensation can be compared in the same time frame.
How does practice change the way a movement is controlled?
Practice rebuilds the model of the forces, and it leaves the shape of the movement alone. Human subjects reaching inside a robot-imposed velocity-dependent force field moved in grossly distorted paths at first, then converged on trajectories close to the ones they drew in free space. Removing the field produced aftereffects that were approximate mirror images of those first attempts, and they appeared in workspace regions the training never visited. That transfer rules out a stored table of visited states and experienced forces.
What does the cerebellum contribute to movement?
The cerebellum carries the prediction error that trains the internal model and then holds the finished model. While people learn a new tool, activity across wide cerebellar areas is proportional to the error signal, and a separate signal near the posterior superior fissure persists after error levels are equalized. Losing that machinery is visible in reaching. Seven adults with cerebellar disease could not produce muscle torques that compensated the interaction torques one moving joint imposes on another, which is what makes their reaches curve, decompose and overshoot.
How do dorsal and ventral premotor cortex differ?
Dorsal and ventral premotor cortex prepare different parts of a reach. Among 211 dorsal premotor neurons recorded in monkeys, 53 percent were tuned to both target location and which arm would be used. Among 109 ventral premotor neurons, 60 percent coded target location alone with no arm selectivity. Inactivation adds a second distinction. Dorsal premotor cortex proves dispensable for visually guided sequences and essential for memory-guided ones, even though comparable numbers of its neurons fire during both tasks. Activity in an area is one measurement and necessity is another.
What does the supplementary motor area do?
The supplementary motor area runs learned sequences of movement without any external cue. In three monkeys performing sequential reaches, more than half of SMA neurons were preferentially or exclusively active for internally guided sequences, 55 percent before movement and 65 percent during it. Premotor neurons favored the visually guided sequences, 55 percent before movement and 64 percent during it. Sequence-specific neurons concentrated in SMA and transition-specific neurons in premotor cortex. Those authors warned that a strict functional dichotomy between the two areas is not acceptable.
Is the motor homunculus accurate?
Within-limb somatotopy in primary motor cortex is neither spatially discrete nor sequentially ordered. Six features constrain it. Output converges and diverges. Horizontal connections spread activity across the map. Activation is distributed, lesions produce characteristic deficits, and the cortex can reorganize. Representations of any two smaller body parts overlap extensively beneath gradual gradients, and outputs to flexor, extensor and synergist muscles spread across one another. The original stimulation data were themselves consistent with distributed organization, so the familiar drawing was over-read.
How does the Unified Model of Tone read motor control?
The Unified Model of Tone reads motor control as a forecast assembled from the body's own report of itself. A segment holding a distorted mechanical state feeds a distorted report into that assembly. The model treats a matched input as information delivered to a state estimate, and not as force applied to a joint, which makes the address the variable under test. In adults with recurrent neck pain, a cervical thrust delivered away from the segment under analysis left the N30 potential where it was, at p equal to 0.757.
How does the neck influence movement of the whole body?
Cervical muscles are built to report position. Counts in human fetal specimens found 190 muscle spindles per gram in obliquus capitis superior, 242 per gram in obliquus capitis inferior, and 98 per gram in rectus capitis posterior, with no tendon organs present. Distort that report and whole-body behavior changes. Blindfolded subjects walking to a target remembered at about 4 meters undershot it and veered toward the side opposite the vibrated muscles. Cervical proprioception sits inside the internal reference the motor system steers by.
11The sources
References
Sources: primary literature, linked inline.