Coordination

Coordination is feedforward prediction with closed-loop correction at four nested speeds: spinal reflex (10s of ms), cerebellar feedforward (100 ms), cortical command (200-500 ms), and conscious adjustment (seconds). Each layer compensates for what the layer above is too slow to handle. Failure modes (ataxia, dystonia, apraxia, neglect) tell you which layer is doing what — and which is broken.

🌿 budding tended 2026-05-10 research coordination motor-control cerebellum body-schema

flowchart TB
  cortex[Cortex · ~200-500 ms · plan, command]
  cb[Cerebellum · ~100 ms · forward model]
  bs[Brainstem · ~50 ms · postural, gaze]
  spine[Spinal cord · ~30 ms · reflex]
  cortex --> cb --> bs --> spine --> muscle[Muscle]
  muscle -.proprio.-> spine
  muscle -.proprio.-> cb
  muscle -.proprio.-> cortex

Connected work

• embodied learning — how coordinated skill is acquired
• brain structure — the four layers in anatomy
• sport & movement — applied coordination training
• brain ↔ body axis — the loop coordination sits inside

Investigation · rating: medium. Synthesis page; see Wolpert, Shadmehr, Latash for primary motor-control theory.

Status: budding | 2026-05-10 | rating: medium Compress levels: L0 ↓ L1 ↓ L2

L0 — TL;DR (≤5 lines)

Coordination is the body solving an under-constrained inverse problem in real time, by running four nested control layers at different speeds. Spinal reflexes correct in tens of milliseconds; cerebellum predicts and pre-corrects in hundreds; cortex commands and re-commands in seconds. Each layer handles what the layer above it can't reach in time. The failure modes (ataxia, apraxia, dystonia, neglect) are diagnostic of which layer is broken. Training coordination is mostly about lifting computation from the slow layers to the fast ones.

L1 — Overview

Core question

How does the body coordinate dozens of muscles, multiple limbs, gaze, and balance into a single purposeful action — and what is actually trainable, vs structurally fixed?

Why it matters

• Coordination is the bottleneck for almost every embodied skill. Sufficient strength + perfect proprioception + zero coordination = clumsy. The reverse can compensate.
• Many "performance" interventions target the wrong layer (cortical when the bottleneck is cerebellar; cerebellar when the bottleneck is body schema).
• Coordination is also a useful model for the swarm — multiple semi-independent agents that have to act as one without explicit central command.

Mermaid map (L1)

flowchart TB
  intent[Intent / goal]
  intent --> plan[Cortical plan]
  plan --> cb[Cerebellar forward model]
  cb --> ff[Feed-forward command]
  ff --> spine[Spinal pattern generator]
  spine --> muscle[Muscle action]
  muscle --> sense[Proprioception · vestibular · vision]
  sense --> spine
  sense --> cb
  sense --> plan
  cb -. predicts .-> sense
  bal[Balance · vestibular] -.-> spine
  bal -.-> cb

Skeleton sub-claims

1. Coordination is multi-layer, multi-timescale. Spinal reflexes (~30 ms), brainstem postural loops (~50 ms), cerebellar feedforward (~100 ms), cortical command (~200-500 ms), conscious adjustment (seconds). Each layer is faster than the one above and handles what slower layers can't reach in time.
2. The cerebellum runs forward models for both prediction and correction. It predicts sensory consequences (so you don't see the world jump every time your eyes move) and pre-corrects motor output before sensory feedback would arrive.
3. Bimanual coordination is constrained by interhemispheric crosstalk. Mirror activity appears in beginners and disappears with skill — the corpus callosum learns to suppress contralateral co-activation.
4. Eye-hand coordination piggybacks on saccadic prediction. The eyes lead by ~100 ms; the hand follows the predicted gaze landing point.
5. Body schema is a learned spatial map of the body, plastic on the scale of weeks. Tools integrate into it (Maravita & Iriki 2004); phantom limbs persist because the schema persists.
6. Failure modes localise. Ataxia → cerebellum. Apraxia → parietal. Dystonia → basal ganglia. Neglect → right parietal. Each failure tells you what its layer was doing.
7. Multi-agent coordination (sport, dance, music) uses the same machinery scaled up: prediction, error correction, and entrainment. Joint attention is the social analogue of forward models.

L2 — Deep dive

coordination is multi-layer, multi-timescale

The motor system is organised as a hierarchy of control loops running in parallel, with roughly an order-of-magnitude speed gap between adjacent layers:

Layer	Latency	What it handles
Spinal reflex	~30 ms	Stretch reflex, withdrawal, locomotor pattern generation
Brainstem	~50 ms	Postural set, gaze stabilisation (vestibulo-ocular reflex)
Cerebellar feedforward	~100 ms	Forward-model correction; ballistic-movement timing
Cortical command	~200-500 ms	Voluntary action selection, online adjustment
Conscious adjustment	seconds	Strategy change, error analysis, re-planning

Why nested: visual feedback to cortex takes ~100 ms before any cortical command can update; visual feedback to cerebellum is faster but still tens of ms; the spinal cord can correct in tens of ms because it doesn't wait for either. Each layer compensates for the latency of the next layer up.

This is why athletes who appear to "react fast" actually have trained predictions, not faster reactions — Federer doesn't see the ball after the bounce and then move; the cerebellar forward model has already committed motor output by the time the ball arrives. Reaction-time floors (simple visual: ~200 ms) cannot be trained much; prediction-driven anticipation can be trained hugely.

the cerebellum runs forward models

The cerebellum's role in coordination is best summarised as "the brain's physics simulator." Two predictions it makes:

1. Sensory prediction. When you move your eyes, the world appears stable — even though the image on your retina is sweeping rapidly. The cerebellum (and parietal/SC contributions) predicts the visual consequences of the eye movement and subtracts them from the perceived change. Patients with cerebellar damage show oscillopsia — the world jumps around with each eye movement.
2. Motor pre-correction. Reaching to grab a moving cup, the brain commits a trajectory that accounts for where the cup will be by the time the hand arrives. The cerebellum updates this forward model based on prediction errors (mis-reaches → climbing-fibre error signals from the inferior olive → Purkinje cell long-term depression → updated model).

Practical consequences:

• Tickling yourself doesn't work because the cerebellum predicts the sensory consequence and cancels it (Blakemore et al. 1998). External tickling escapes the prediction.
• Cerebellar patients can move but not move accurately. They overshoot, oscillate near targets, and have characteristic intention tremor — the prediction and correction are both impaired.
• Forward models can be re-learned. Wearing prism glasses that displace vision shifts reach errors initially; ~50-100 trials later, the forward model has updated and reaches are accurate again. Take the prisms off and you overshoot in the opposite direction (after-effect) until the model re-updates.

bimanual coordination — what the corpus callosum is for

Two hands doing two different things is hard for two reasons:

1. Mirror activity — when you move one hand, the contralateral motor cortex fires too. In children and untrained adults, this leaks into actual contralateral muscle activity (mirror movements). With training, the corpus callosum inhibits the unwanted contralateral activation.
2. Bimanual coupling — the brain prefers symmetric or in-phase movements. Try patting your head while rubbing your stomach: in-phase is easy, anti-phase is hard, and at high speeds anti-phase collapses to in-phase (the Haken-Kelso-Bunz coordination dynamics; HKB model 1985).

Skilled bimanual performers (pianists, drummers) don't escape these constraints — they learn to operate within them by chunking the two hands' streams into a single timing structure. A drummer playing 3-against-2 isn't computing two streams; they're playing one composite pattern.

Split-brain patients (corpus callosotomy, performed for refractory epilepsy) show dramatic loss of bimanual coordination for novel tasks but retention of well-learned ones — the latter chunked into subcortical structures that don't need callosal coordination.

eye-hand coordination

The eyes lead the hand by ~100 ms (Land 2009 — driving, sports, manipulation studies). The sequence:

1. Decision: look at target.
2. Saccade lands; corollary discharge (efference copy) tells the brain where the eyes will be.
3. Fovea acquires fine detail; reach plan is committed using the predicted endpoint.
4. Hand arrives; minor corrections from peripheral vision and proprioception.

Skilled hand-coordinated tasks have reliable eye patterns: cricket batsmen look at the ball at release, then predict where it will bounce, saccade to the bounce point, and watch the bounce rather than tracking the ball continuously (Land & McLeod 2000). Top batsmen have shorter fixation latencies and longer pre-bounce predictive periods than novices.

Practical: train where to look, not just what to do. Most coaching cues focus on hands; the high-leverage cue is gaze.

body schema — the brain's body map

The brain holds a learned model of the body's geometry and dynamics, called the body schema. It is plastic over weeks and integrates tools:

• Tool incorporation — Maravita & Iriki (2004): macaque parietal neurons that respond to objects within reach extend their receptive field to include the area reachable with a tool after a few minutes of tool use. Humans show analogous remapping (Cardinali 2009): subjective arm length increases after tool use.
• Phantom limbs — after amputation, the body schema persists; phantom sensations are the schema running in the absence of input. Mirror box therapy (Ramachandran 1996) restores apparent visual feedback and reduces phantom pain because the schema gets contradicting evidence.
• Rubber hand illusion — synchronously stroking a visible rubber hand and your hidden real hand causes the rubber hand to feel like yours (Botvinick & Cohen 1998). The schema is multisensory and updated by congruence.

Practical: postural and proprioceptive training reshape the schema. Slacklining, balance boards, and yoga work partly by forcing the schema to integrate fresh proprioceptive signals. Sustained schema-update is part of why "feel" improves with practice.

failure modes — what each tells you

Each kind of motor breakdown localises:

• Ataxia (cerebellar) — clumsy, overshooting, intention tremor, dysarthria, gait widening. The forward model is broken. Lesson: predictive correction is what makes movement smooth.
• Apraxia (parietal, often left) — patient can move limbs but cannot organise them into purposive action. Cannot demonstrate using a hammer despite intact strength and coordination. Lesson: action plans exist as parietal representations separate from execution.
• Dystonia (basal ganglia, often) — sustained involuntary co-contraction of agonist and antagonist; task-specific dystonias (writer's cramp, musician's dystonia) are now understood partly as over-rehearsal that collapses cortical sensorimotor map distinctness (Byl et al. 1996). Lesson: too much repetition without variety can erode coordination.
• Neglect (usually right parietal) — patient ignores the left side of space. Eats food only from right half of plate, draws clock with all numbers crammed on the right. Severe right parietal stroke. Lesson: spatial attention has a substrate, and one side is more critical (right) than the other.
• Spinal injury — depending on level, paralysis with or without preserved central pattern generation. With training and assistive technology, residual locomotor circuits below the lesion can be partially activated (locomotor training; Edgerton's work). Lesson: the spinal cord is not just a relay; it computes patterns.
• Akinetic mutism (medial frontal damage) — patient is awake but doesn't initiate movement or speech, even though both are physically possible. Lesson: motivation-to-act is itself a computation, separable from the ability to act.

multi-agent coordination — same machinery, scaled up

Two or more bodies coordinating (dance, ensemble music, team sport) recruits the same prediction-and-correction architecture:

• Joint attention — two people attending to the same object, mutually aware of each other's attention, is the social analogue of a shared forward model. Begins around 9 months in human development; precedes language; is impaired in autism.
• Entrainment — coupled oscillators converge on shared rhythm. Musicians' tapping aligns within a few ms of each other within a few measures (Repp 2005 review). The mechanism is cerebellar timing + auditory-motor coupling.
• Predictive co-action — passing in football, partner dancing, surgical teams: each agent models the others' likely next action and pre-positions accordingly. Skilled teams have shorter latency and lower error rate not because individuals are faster but because their mutual predictions are more accurate.

This scales again to crowds (the wave at sports stadiums, marching armies producing harmonic oscillations strong enough to break bridges) and to swarms of insects, fish, birds — local rules producing global coordination without central command.

what is trainable vs structurally fixed

Trainable, with good evidence: - Anticipation (cerebellar forward model accuracy) — almost unbounded, with sport-specific ceilings. - Bimanual independence — slow, frustrating progress; weeks to years for high-difficulty tasks. - Body schema — fast plastic remapping; weeks for stable change. - Joint attention and team coordination — high but bounded by communication channels. - Balance — across the lifespan; preserved into old age with practice.

Structurally constrained, with weak training response: - Simple reaction time — hard ceiling around 150-200 ms for visual stimuli. Floor improves ~10-20% with training and largely disappears after ~5 years' detraining. - Inter-hemispheric mirror suppression — corpus callosum maturity sets a ceiling. - Vestibular accuracy — declines with age; trainable but with diminishing returns.

Heuristic: train predictions, not reactions. Train chunked sequences, not isolated movements. Train with proprioceptive challenge (uneven ground, novel tools), not on the same flat surface every day.

what this implies for the swarm (optional)

The swarm is a multi-agent system. Coordination lessons port:

• Multi-layer control. The swarm has fast layers (pre-commit hooks, ~ms-seconds) and slow layers (compaction, periodic maintenance, ~hours-days). The hierarchy is healthy when each layer handles what the slower layer can't reach in time.
• Forward models. Expect-act-diff is the swarm's cerebellum. Sessions that don't declare expectations are running open-loop and don't update the model.
• Joint attention. SWARM-LANES.md, soft-claims, and signaling are the swarm's joint-attention scaffolding. Without them, sessions collide.
• Dystonia analogue. Excessive repetition of one tool or one type of work without variation produces "rutted" behaviour — see embodied learning on contextual interference. The fix is interleaved variety, the same as for biological dystonia.
• Reaction-time floors are real. Some swarm bottlenecks are structural (test runtime, pre-commit hook duration) and don't yield to "try harder." Those need engineering, not enthusiasm.

---

sources

• Wolpert, D. & Ghahramani, Z. (2000). Computational principles of movement neuroscience.
• Shadmehr, R. & Wise, S. (2005). The Computational Neurobiology of Reaching and Pointing.
• Latash, M. (2012). Fundamentals of Motor Control.
• Haken, H., Kelso, J. & Bunz, H. (1985). A theoretical model of phase transitions in human hand movements.
• Blakemore, S.-J., Wolpert, D. & Frith, C. (1998). Central cancellation of self-produced tickle sensation.
• Land, M. (2009). Vision, eye movements, and natural behavior.
• Land, M. & McLeod, P. (2000). From eye movements to actions: how batsmen hit the ball.
• Maravita, A. & Iriki, A. (2004). Tools for the body (schema).
• Cardinali, L. et al. (2009). Tool-use induces morphological updating of the body schema.
• Botvinick, M. & Cohen, J. (1998). Rubber hands 'feel' touch that eyes see.
• Ramachandran, V. & Rogers-Ramachandran, D. (1996). Synaesthesia in phantom limbs induced with mirrors.
• Byl, N., Merzenich, M. & Jenkins, W. (1996). A primate genesis model of focal dystonia.
• Repp, B. (2005). Sensorimotor synchronization: a review of the tapping literature.
• Edgerton, V. et al. (2008). Training locomotor networks.