Motor Control III: The Remarkable Brain
We’ve learned how the brain controls the golf swing and how it learns. Now we step back and ask: how remarkable is what the brain actually accomplishes?
When you swing a golf club, your brain is simultaneously:
- Controlling roughly 200 muscles across 20+ joints
- Working with 30–50 millisecond neural delays
- Processing 80–150 millisecond sensory delays
- Generating force within 50–100 millisecond muscle activation delays
- Approximating something like a nonlinear optimal control problem in real time
- Managing transitions between high-control and high-drift phases
- Doing all of this in 300 milliseconds with only 20 watts of power
As of this writing, no widely deployed robot or AI system matches this exact combination of speed, dimensionality, adaptability, and low-power execution in an everyday athletic setting. Yet your brain does something close to it every day, without conscious calculation. This is the chapter that asks: how? And what does this remarkable biological achievement suggest about the nature of intelligence?
We also ask the inverse question: why does the brain occasionally fail catastrophically? Why do the yips happen? Why do skilled golfers sometimes choke? The answers reveal surprising truths about how the brain works.
Finally, we consider what artificial intelligence could learn from the brain, and what the golf swing reveals about the future of AI.
The Scale of the Problem the Brain Solves
Let’s quantify exactly what the brain accomplishes in a 300 millisecond downswing.
The Dimensionality of the System
Degrees of freedom: The golf swing involves roughly 7 major joints: shoulders (2 DOF each = 4), elbows (1 each = 2), wrists (2 each = 4), torso rotation (1), hip rotation (1), knee bend (1), ankle (1). Total: 13 DOF for major joints. Including smaller joints (fingers, feet) and the club as a separate dynamic object: 20+ DOF.
Muscles: Roughly 200 muscles are involved in the swing, from large muscles (pectoralis, latissimus dorsi) to small stabilizer muscles. Each muscle can be activated at a different level, producing a huge control space. As a thought experiment, even discretizing each muscle to just 7 activation levels gives \(7^{200} \approx 10^{169}\) states—a number so large it exceeds the number of atoms in the observable universe (\(\sim 10^{80}\)). The point is not that the brain literally searches this space; the point is that structure must drastically reduce the effective control problem.
Time points during the swing: At 100 Hz temporal resolution (10 milliseconds per time step, reasonable for neural control), a 300 ms downswing has 30 time points. At each time point, you could independently command each muscle.
Control space size: If you could independently vary muscle activations at each time point, the total number of possible swings would be roughly \((7^{200})^{30}\). This thought experiment produces an astronomically large space: \(\approx 10^{4000}\) distinct swings.
Computational challenge: To find the optimal swing by exhaustive search would require evaluating \(10^{4000}\) candidates. Even if each evaluation took a nanosecond, this would take \(10^{3991}\) seconds—far longer than the age of the universe.
Yet the brain produces workable solutions in milliseconds. How?
Constraints Reduce the Space
The brain doesn’t solve this high-dimensional problem directly. Instead, it exploits structure in the problem. Several constraints dramatically reduce the effective dimensionality:
Synergies: Instead of controlling 200 muscles independently, the brain activates muscle synergies—fixed patterns of muscle co-activation. Within a given movement class (e.g., reaching tasks), synergies account for roughly 80–90% of the variance in muscle activations (Bizzi et al. 2008; Cheung et al. 2005). Across diverse movement types, the explained variance drops to 60–75%, suggesting that synergies are context-dependent building blocks rather than universal motor primitives.
A synergy might be: “activate shoulder internal rotator + elbow extensor + wrist extensor,” which produces a synergistic motion. By reducing from 200 independent muscles to 8 synergies, the dimensionality drops from \(7^{200}\) to \(7^8 \approx 5.7\) million.
Temporal structure: The brain doesn’t command each of the 30 time points independently. Instead, it specifies a trajectory—a smooth function of time. Smooth functions can be represented with fewer parameters. For example, a cubic polynomial has 4 parameters but can describe a complex trajectory.
By using smooth trajectories for each synergy, the control dimensionality drops to perhaps 30–50 parameters.
Physics constraints: The dynamics are highly nonlinear, but they have structure. In the affine-control perspective developed in this book, late-swing drift can become large relative to newly applied control inputs. If that picture is approximately right, then many nearby motor commands may lead to similar outcomes, allowing the brain to rely more on state preparation and less on continuous correction.
Learned motor primitives: The brain doesn’t solve the problem from scratch each swing. It retrieves a learned primitive—a pre-computed motor program—and applies it. This is like having a solution template that only needs minor tuning.
These constraints reduce the effective dimensionality from \(10^{4000}\) to something manageable: perhaps a few dozen parameters to tune, though the exact number is heuristic rather than literal.
The Constraint of Time
The most severe constraint is temporal:
The downswing lasts 300 milliseconds.
Neural transmission from brain to muscle: 30–50 milliseconds.
Muscle activation delay: 50–100 milliseconds.
Sensory feedback delay: 30–150 milliseconds (proprioception is fast, vision is slow).
By the time visual feedback about the swing reaches your brain, the swing is over. By the time proprioceptive feedback produces a corrective motor command, the swing is over.
The brain can likely process only a very small number of meaningful feedback correction cycles during the downswing.
Conclusion: the downswing is likely to be heavily pre-planned, with limited opportunity for deliberate real-time correction.
This is not a limitation of the brain; it is a fundamental constraint of the system. Even with very fast neural processing, real-time feedback control of the downswing would be limited by sensory and actuator delays. The practical solution is to rely heavily on feedforward structure, while still allowing fast reflexive and stabilizing feedback to contribute where it can. Optimal-control accounts of sensorimotor behaviour reach the same place from the other direction: what correction the nervous system does apply is concentrated on the dimensions that affect the task and withheld elsewhere, so a movement can be both tightly controlled where it matters and largely uncorrected everywhere else (Todorov 2004).
How the Brain Manages: Strategies for a Bandwidth-Limited Controller
Given these constraints, how does the brain manage to swing a golf club?
Strategy 1: Exploit Drift—Use Physics as the Primary Actuator
This is the central insight of the textbook, viewed from the motor control perspective.
In the affine-control interpretation developed in this book, the golf swing can become drift-dominated late in the downswing. In that view, the state-dependent mechanical terms (gravity, inertial effects, elastic restoring forces) can become large relative to newly applied control torques.
If that interpretation is correct, it gives the brain an enormous advantage. The brain does not need to generate the whole trajectory from scratch. It mainly needs to:
- Set the initial conditions correctly (address position, grip orientation, muscle stiffness).
- Apply an early-swing control input that initiates the acceleration.
- Let the drift field carry much of the system through the mid- and late-swing phases.
Shot outcome is then shaped strongly by the evolving system state and passive dynamics, not only by precise real-time muscle control. In that case, the brain can exploit a high-level setpoint control strategy: specify where the system should be at key points, and let physics carry much of the motion between them.
This is why the ZTCF family framework (Chapter 6) is so important. It offers a way to ask whether a seemingly complicated motion contains a large state-driven component that reduces the burden on moment-to-moment control.
Compare this to trying to control a tennis racket during a return, where the control forces may be more comparable to drift forces and the time window is even shorter. This would require more sophisticated real-time feedback control. The hypothesis is that the drift-dominated structure of the golf swing is part of what makes learned consistency possible, though this comparison has not been rigorously tested across sports.
Strategy 2: Impedance Control—Command Stiffness, Not Position
A second strategy the brain uses is impedance control. Instead of commanding exact joint angles or forces, the brain commands stiffness.
The resistance of a joint to perturbation. A very stiff joint (high co-contraction of agonist and antagonist muscles) is hard to move and resists external forces. A compliant joint (low co-contraction) is easy to move and absorbs forces.
Instead of the brain commanding: “Move the shoulder to this exact angle at this exact rate,” it may command something closer to: “Set shoulder stiffness to a task-appropriate level” (for empirical ranges from impedance-control experiments, see (Hogan 1985b)).
This is a partly passive control strategy. The stiffness is a mechanical property of the musculoskeletal system once it has been established through muscle activity. After it is set, the joint naturally resists perturbations without requiring continuous active correction.
Mathematically, impedance control commands the effective stiffness \(\bm{K}\) and damping \(\bm{D}\): \[ \bm{u} = -\bm{K}(\bm{q} - \bm{q}_d) - \bm{D}\dot{\bm{q}} \]
where \(\bm{q}_d\) is a desired (but “soft”) target position. Small deviations from \(\bm{q}_d\) produce restoring torques. Large deviations don’t.
This is much simpler than controlling exact position because the stiffness is set once, not recomputed at each time step.
For the golf swing, impedance control explains several phenomena:
Grip pressure: A light grip (low impedance) allows wrist hinge freedom but makes the club unstable. A firm grip (high impedance) locks the wrist but limits power. Great golfers learn the optimal grip stiffness—high enough to stabilize the club, low enough to allow hinge.
Address posture: Setting up with relaxed muscles (low impedance) vs. rigid muscles (high impedance) changes how the swing unfolds. Low impedance is generally better because it allows the system to find optimal paths. Too-rigid posture constrains motion.
Reactive responses: If the ball suddenly moves (unlikely in golf, but imagine it), or if wind perturbs the swing, a joint with higher impedance resists the perturbation more. The golfer doesn’t need to consciously correct; the mechanical impedance does it.
Load distribution at impact: High-impedance impact absorption spreads forces across multiple joints, which biomechanical models suggest may help distribute peak loads at impact.
Setting impedance is more efficient than precise position control because it doesn’t require fast computation. The brain can set impedance once and let mechanics handle the rest (Hogan 1985a; Mussa-Ivaldi and Bizzi 2000).
Strategy 3: Synergies—Reduce Dimensionality Through Coordination
We mentioned synergies earlier. Let’s examine them in detail.
A fixed pattern of muscle activations that produces a coordinated action. Instead of independently controlling each muscle, the brain activates predefined synergy vectors.
Synergy 1 (pulling): “Activate latissimus dorsi + pectoralis + internal shoulder rotators.” This produces a pulling motion—useful in the downswing.
Synergy 2 (pushing): “Activate deltoid (posterior) + triceps + wrist extensors.” This produces a pushing/extending motion.
Synergy 3 (stabilizing): “Co-activate all shoulder muscles equally.” This stabilizes the shoulder without moving it.
By combining these 3–5 synergies at different activation levels across time, the brain produces the full complexity of the golf swing.
The advantage: instead of commanding 200 muscles independently, the brain commands 4–5 synergies. Dimensionality drops dramatically.
Evidence for synergies comes from electromyography (EMG) recordings of muscle activation during movement. Researchers have found that most of the variance in muscle activations across many different movements can be explained by 4–8 synergy vectors (Bizzi et al. 2008; Cheung et al. 2005).
Why would synergies evolve? One hypothesis: they simplify learning. Instead of learning a 200-dimensional control space, the brain learns a 5-dimensional space. This is much easier and happens faster (Bizzi et al. 2008; Cheung et al. 2005; Torres 2013).
Strategy 4: Predictive Control and Anticipatory Activation
The brain uses its internal model to predict what will happen and pre-activate stabilizing muscles.
Suppose you know that lifting a heavy object will create a sudden force. Instead of waiting for the force and reacting, the brain pre-activates core muscles before the perturbation. This is anticipatory control.
For the golf swing:
The transition is coming. The brain pre-activates muscles that will resist the sudden acceleration.
The clubhead will soon hit the ball. The brain pre-activates muscles that will stabilize impact.
These pre-activations happen tens of milliseconds before the event, anticipating it using the forward model.
The numerical values, claims, and parameters discussed in this section are formulated as illustrative model outputs and didactic simplifications. They are intended for pedagogical purposes and do not represent published, empirical biomechanical measurements.
- In this illustrative anticipatory sequence, by pre-positioning muscles, the brain avoids the neural delays. The muscles are already primed when the event occurs Wolpert and Kawato (1998).
This is why the pre-swing motion and address posture are so critical. Elite golfers often spend several seconds setting up precisely. They’re not just getting comfortable; they’re pre-positioning the musculoskeletal system and pre-activating muscles. The actual swing, when it comes, unfolds from a carefully prepared state.
The Brain as an Internal Model: What It Learns
Chapter 24 and Chapter 25 established that the brain learns internal models. Let’s now think carefully about what, exactly, the brain learns.
Through practice, the brain learns three things about the golf swing:
The drift field \(\hat{f}(\bm{x})\): What physics does for free. Which parts of the trajectory are gravity-driven, which are active. Where the trajectory curves naturally.
Evidence: when you change clubs (different weight, different mass distribution), your swing is bad for the first few attempts. The internal model for the drift field is wrong for the new club. After a few dozen swings, your internal model updates, and performance recovers.
The control effectiveness \(\hat{G}(\bm{x})\): What muscles can accomplish. How much force each muscle can produce, how long it takes to produce it, which joints are coupled.
Evidence: a beginning swimmer thrashes ineffectively; muscles are firing in opposition. An elite swimmer flows efficiently; muscles fire in synergy. The internal model of \(\hat{G}(\bm{x})\) is much more accurate.
The optimal policy \(\bm{u}^*(\bm{x})\): What to do at each state to reach the goal. This is the learned motor program.
Evidence: after thousands of swings, your golf swing becomes automatic. You’ve learned the policy—given any state during the swing, fire these muscles.
Learning the drift field IS learning physics. The brain becomes a physicist, even if it doesn’t use equations. The neural encoding in the cerebellum and motor cortex approximates the nonlinear dynamics of the golf swing.
How Do We Know the Brain Learns Physics?
Several lines of evidence show that the brain learns a model of physics:
Experiment 1: Novel tool learning. A golfer learns to swing a driver for years. Then they pick up a putter. The putter is dramatically different: much shorter, much lighter, different mass distribution. The golfer’s first few putts are terrible—they overshoot, undershoot, or miss entirely.
Within 10–20 putts, performance recovers. The internal model has updated for the new tool dynamics.
Within 100 putts, the golfer has learned the putter physics. They can hit putts of different distances, on sloping greens, etc.
If the brain didn’t learn physics, this wouldn’t happen. The brain would have no way to generalize from driver swings to putter strokes. But it does generalize, by updating its internal model.
Experiment 2: Perturbed feedback. Researchers study learning in a 2D reaching task where the visual feedback is systematically rotated. For example, when you try to move your hand right, the feedback shows it moving at a 30-degree angle.
Initially, people make large errors. But within 30–40 trials, they adapt. Their reaching becomes accurate again.
Critically, when the perturbation is removed, performance is initially poor (negative aftereffects), then recovers. This shows that the brain has internally updated its model to compensate for the perturbation.
This adaptation is too fast and too precise to be explained by conscious trial-and-error. It’s the cerebellum updating the forward model.
Experiment 3: Generalization to novel contexts. After adapting to a rotated feedback perturbation in one movement direction, the brain partially adapts to other movement directions, without explicit practice in those directions.
Why? The internal model has learned something about the general transformation—not just memorized the specific case. This is learning the underlying physics, not memorizing a response (Wolpert et al. 2011; Shadmehr and Mussa-Ivaldi 2012).
Comparison With Artificial Intelligence: What Can Machines Learn?
The brain is an existence proof that biological neural networks can learn to control the golf swing. Can artificial neural networks do the same?
How Brain Learning Differs From AI Learning
| Aspect | Brain | AI (Reinforcement Learning) |
|---|---|---|
| Data efficiency | 10,000–100,000 trials | \(10^6\)–\(10^8\) simulated episodes |
| Learning time | Hours to years | Seconds to hours (simulated) |
| Sensory input | Rich (proprioception, vision, touch, vestibular) | Impoverished (usually joint angles only) |
| Built-in structure | Yes (cerebellar circuit, motor cortex hierarchy) | Often not (generic architecture) |
| Error signals | Multiple (proprioceptive error, outcome error, dopamine) | Usually single reward signal |
| Learning rule | Local (synaptic level) | Global (backpropagation) |
| Online vs. offline | Both | Typically offline |
| Interpretability | Some (we understand cerebellar circuit) | Poor (black box) |
The brain is sample-efficient. It learns from roughly 10,000 repetitions. Reinforcement learning algorithms often need \(10^6\) to \(10^8\) simulated episodes. Why?
Several factors:
Richer feedback: The brain gets proprioceptive feedback at every moment (prediction errors on the trajectory), not just a final reward. This means more learning signals per trial.
Built-in structure: The brain’s neural circuits are not random. The cerebellum has a stereotyped architecture that’s good for learning dynamics. This prior structure accelerates learning.
Transfer learning: The brain leverages prior knowledge from other movements. A tennis player learning golf already has some internal models for rotational motion. An AI starting from random weights has no prior knowledge.
Hierarchical control: The brain decomposes the problem. High-level planning, mid-level trajectory generation, low-level muscle control. This decomposition makes learning at each level easier.
What the Brain Does Better
Transfer learning: Humans transfer skills across very different contexts (tennis to golf, squash to badminton). AI systems are often brittle and task-specific.
Robustness: Humans handle perturbations gracefully. If the wind blows during your swing, you adjust. If you’re tired, you adapt. AI systems often fail on out-of-distribution inputs.
Compositionality: Humans combine learned skills in novel ways. A golfer who’s played tennis and baseball might discover a unique swing combining elements of both. AI systems rarely compose skills.
One-shot learning: Humans can sometimes learn from a single demonstration (“Do it this way”). Deep RL typically needs thousands of examples.
Physical understanding: Humans develop intuitive physics through play and exploration. “That club is heavier, so I need to swing slower.” AI systems don’t naturally develop this understanding.
What AI Does Better
Consistency: A trained neural network produces the same output for the same input, every time. The brain sometimes gets the yips.
Speed: Modern AI can evaluate millions of candidates per second. The brain processes at 10–40 Hz.
Global optimization: AI can find globally optimal solutions using large-scale search. The brain may find local optima. (Though in practice, local optima are often good enough.)
Scalability: AI can be trained on massive datasets. The brain’s learning is limited by the time humans can spend practicing.
Precise control: AI can command exact forces with millisecond precision. The brain has noise and delays.
What Artificial Intelligence Could Learn From the Brain
If the brain is so impressive at learning motor control, what could AI learn from it?
Hierarchical Control
The brain doesn’t try to solve one massive optimization problem. Instead, it decomposes:
Strategic level (prefrontal cortex): Choose goal. “I want to hit my target 150 yards away with a smooth swing.”
Tactical level (premotor cortex): Plan trajectory. “I need the club to reach maximum speed at impact, with face angle matching target line.”
Execution level (motor cortex + cerebellum + spinal cord): Generate commands. “Fire these muscles in this sequence.”
This hierarchy reduces the problem at each level. The strategic level doesn’t think about muscles; the execution level doesn’t think about goals.
Modern AI research has begun exploring similar hierarchies. The options framework in reinforcement learning formalizes this: high-level actions (options) are themselves learned policies for lower-level control (Barto 2003; Sutton and Barto 2018).
Internal Models and Model-Based RL
The brain builds an explicit forward model of the world. Modern AI has been dominated by model-free RL (like Q-learning), which learns value functions without learning a model.
But there’s growing recognition that model-based RL is more sample-efficient. If an agent has a forward model, it can imagine many trajectories without actually executing them.
Model-free approach: The robot tries a motor command. It gets a reward. It updates its value function. It tries another command. After a million trials, it learns which commands are good.
Model-based approach: The robot learns a forward model: “If I move in this direction with this force, I’ll end up here.” Now it can imagine many trajectories using the model, without physically executing them. It can plan offline and execute online.
The brain uses model-based approach. Modern AI is moving toward it (Deisenroth et al. 2015; Hafner et al. 2020).
Embodiment and Physics Exploitation
A profound insight from the brain is that it doesn’t try to control every DOF. Instead, it exploits the body’s morphology and the environment’s physics.
The golf swing works because gravity does most of the work. The brain doesn’t fight gravity; it uses it.
More generally, the brain exploits:
Passive dynamics: The body’s natural resonance frequencies and elasticity. Swinging your arm is easy because the arm naturally oscillates at a certain frequency. Swing at that frequency and forces are minimized.
Environmental affordances: Properties of the environment that enable action. The golf ball’s elasticity, the club’s elasticity, the ground’s firmness—all of these enable efficient control.
Morphological computation: The shape of the body itself computes control. A limb with elastic tendons and springy muscles stores and releases energy with minimal neural control.
Most roboticists try to control everything actively. A better approach: design the robot to exploit passive dynamics, and use active control only for the fine adjustments (Ijspeert 2014; Ghazi-Zahedi and Ay 2013).
Prediction-Driven Learning
The brain learns from prediction errors. At every moment during the swing, the brain predicts what proprioceptive signals it expects, compares to actual signals, and learns from the mismatch.
This is self-supervised learning: the agent generates its own learning signal by predicting its own sensory feedback.
Modern AI has begun exploring self-supervised learning. Instead of learning from hand-labeled data (expensive) or reward signals (sparse), agents learn by predicting future observations.
Traditional approach: Train a robot arm on 1000 manually labeled examples of “reaching to target.”
Self-supervised approach: Let the robot arm move randomly for an hour. For each action, predict the resulting arm configuration. Learn a forward model by predicting its own sensory consequences. Now the robot can use this forward model for planning without any labeled data.
This is closer to how the brain learns. A baby doesn’t have labeled data. Instead, it acts, observes consequences, and builds a forward model of its body (Pathak et al. 2017; Hafner et al. 2020).
The Golf Swing as a Window Into Intelligence
The golf swing is a perfect test case for understanding intelligence. Let’s see why.
Why Golf Is a Useful Scientific Tool
Complexity: Golf involves dozens of joints, hundreds of muscles, nonlinear dynamics, and high-dimensional optimization. It’s complex enough to be challenging.
Simplicity: Unlike real-world tasks (catching a ball while running), golf is highly controlled. The ball doesn’t move. The environment is constant. This simplicity makes analysis tractable.
Measurability: The outcome is objective and easy to measure: where did the ball land? How far? The consistency of expert performance can be quantified precisely.
Universal expertise: Golf is learned and practiced by millions. There’s enormous variability in skill levels, from novices to world champions. This variation provides a natural experiment in learning and expertise.
Introspection: Golfers are highly self-aware about their performance. They can report on what they feel, what they think about, what works and what doesn’t. This introspection is valuable data.
Failure modes: Golf has interesting failure modes (slumps, yips, choking) that reveal how the nervous system works. Understanding why a skilled player suddenly can’t hit a 3-footer illuminates nervous system function.
Timescale: Learning golf takes years. This allows studying learning over long timescales that are hard to access in the lab.
Robustness testing: Golf tests robustness to perturbations. Wind, uneven lies, fatigue, emotion, pressure—all degrade performance. A comprehensive theory of motor control must explain robustness.
Several research groups have used golf to study motor control. Examples:
Vickers (quiet eye): Studied gaze fixation in expert golfers, showing that the quiet eye (prolonged fixation on the ball) predicts performance.
Schöllhorn (differential learning): Used golf to study whether practice variability improves learning better than blocked practice.
Bayes (swing biomechanics): Developed kinematic models of the golf swing to understand inter-individual variability.
McLennan (neurocognitive factors): Studied cortical activation during golf using fMRI to understand how the brain changes with expertise.
Open Questions in Motor Control That Golf Helps Address
How is the drift field represented neurally? The brain clearly has an internal model of physics. But how? What neural population codes encode the drift field? How does this encoding change with learning?
How does the brain choose among redundant solutions? Given that there are many ways to swing a club (many motor programs that produce good outcomes), how does the brain choose one? Is it minimum effort? Minimum variance? Something else?
What causes the yips? Why do skilled golfers suddenly lose the ability to hit 3-footers? Is it a corrupted motor program? Excessive conscious attention? Fear-induced changes in the cerebellum?
How do experts avoid choking under pressure? Under pressure, performance sometimes degrades (choking). But expert golfers are more resistant to choking. Why? Is it deeper motor memory? Better emotion regulation? Automatic execution that bypasses conscious interference?
Can we use brain imaging to see the internal model? With fMRI or other neuroimaging, can we decode what forward model the brain has learned? Can we see differences between experts and novices in their neural representations?
How is motor learning consolidated? Learning progresses from highly variable (cognitive stage) to highly consistent (autonomous stage). What neural changes underlie this consolidation? How does sleep affect consolidation?
What is the nature of motor skill transfer? Why do tennis players learn golf faster? What knowledge transfers? Is it the internal model of rotational dynamics? The sense of timing?
These are deep questions in neuroscience and motor control. Each has implications for understanding intelligence itself.
Why the Brain Sometimes Fails: The Remarkable Failures
The brain solves the golf swing problem remarkably well, most of the time. But sometimes it fails spectacularly. These failures are informative.
The Yips: A Corrupted Motor Program
The yips are a condition where a golfer loses the ability to perform a previously automatic task. Typically, it strikes short putting (4-footers and closer) but can affect other aspects of the swing.
A golfer with the yips will address a 3-footer that they’ve made hundreds of times before. They address the ball. They begin the stroke. Their hands suddenly jerk involuntarily. The ball skids left or right or backward.
A sudden, involuntary disruption of motor control in a previously learned, automatic movement. The disruption is specific to the task (golfers with putting yips can hit drives fine). It worsens under pressure. It often persists for years.
What causes the yips?
Motor program corruption: The reference trajectory stored in the cerebellum has become corrupted or unstable. Each time the golfer tries the motion, they get a different proprioceptive feedback pattern. This mismatch produces an error signal that can’t be resolved. The result is jerky, unstable motion as the motor system tries to correct continuous errors.
Excessive conscious attention: The golfer is thinking about mechanics instead of feel. This overthinking bypasses the automatic motor system and engages the explicit, conscious system. The conscious system is slow and unstable for fast movements. Result: yips.
Fear-induced changes: Repeated failures (missing putts) produce stress and fear. The amygdala becomes hyperactive. Fear signals modulate cerebellar function, impairing the normal learning and stability of motor programs.
Focal dystonia: The yips might be a form of focal dystonia—an involuntary muscle contraction disorder. This is a neurological condition (disorder of basal ganglia and motor cortex) rather than a psychological one.
The yips teach us that motor control is fragile. Once you’ve learned a movement, you can’t easily unlearn it or consciously control it. The more automatic the movement, the harder it is to consciously correct.
This has a profound implication: overanalyzing your swing is dangerous. The more you think about mechanics, the more you disrupt the automatic system. Paradoxically, the best way to improve is often to stop thinking and just feel (Beilock 2010; Gallwey 1998).
Choking Under Pressure
Another interesting failure mode is choking—a sudden, temporary decrement in performance under high pressure.
You’re 1-down in a match. You have an 8-footer to win. You address the ball. Your heart is pounding. Your hands are shaking. You swing and miss the putt. You lose the match.
Later, on the range, you hit that same putt 9 times out of 10.
What happened? Under pressure, something changed in how your brain controlled the movement.
Choking is well-studied in psychology. The leading theory: conscious processing hypothesis (Beilock 2010).
Normally, skilled movements are automatic. The golf swing unfolds without conscious attention. The golfer thinks about the target, not the mechanics.
Under pressure, the golfer becomes anxious. Anxiety triggers conscious, explicit attention. Instead of automatically swinging, the golfer starts thinking about mechanics: “Keep your head still. Smooth tempo. Follow through.”
But explicit, conscious control is slow and imprecise. The autonomous system, which is fast and precise, is disrupted by the explicit system trying to micromanage.
The result: the normally smooth, automatic swing becomes jerky and error-prone.
This is called the processing trade-off: expertise involves automaticity. Conscious attention disrupts automaticity. Under pressure, conscious attention increases, so automaticity is disrupted.
How do experts resist choking? Several strategies:
Pre-performance routines: A structured routine (deep breaths, target selection, waggle) keeps conscious attention directed at the task goal, not mechanics. The routine is automatic, so it doesn’t disrupt the swing.
Attention control: Training to direct attention externally (on the target) rather than internally (on mechanics). This is ideomotor theory in practice.
Stress inoculation: Practice under pressure conditions. High-level golfers practice in tournaments, with gallery watching, with real stakes. This trains the nervous system to perform under stress.
Memory of success: Recalling previous success under pressure. “I’ve made this putt in high-pressure situations before. I can do it.” This activates the learned automatic program.
Reappraisal: Reframing pressure as excitement. “My heart is pounding because I’m excited and ready,” not “because I’m nervous and afraid.” This changes how the amygdala modulates motor control.
These observations have led to coaching approaches that emphasize routine, external focus, and stress management (Beilock 2010; Baumeister 1984).
The Humbling Conclusion
Let’s return to the central observation. Every time you swing a golf club, your brain accomplishes the following in 300 milliseconds:
Solves a 20+ DOF nonlinear optimal control problem.
Works with 30–50 millisecond neural delays.
Receives feedback delayed by 30–150 milliseconds, too late for feedback correction.
Manages the transition from high-control (high-impedance, slow) to high-drift (low-impedance, fast) regimes.
Coordinates 200 muscles across dozens of joints simultaneously.
Maintains consistency: elite golfers produce clubhead speeds within 1–3% of their average, swing after swing (Jorgensen 1994).
Exploits physics: uses the drift field (gravity, centrifugal effects) to do most of the work, rather than fighting physics.
Uses minimal power: accomplishes all of this with 20 watts of energy.
Does this automatically, without conscious attention, often while thinking about something completely different.
As of this writing, no robot matches this combination of dimensionality, adaptability, and energy efficiency. No AI system has demonstrated comparable performance in embodied dexterity tasks of this complexity, though the field is advancing rapidly.
And yet, the brain shanks.
This is the deepest truth about the golf swing. It is, simultaneously, the most impressive control task your brain ever accomplishes and the easiest thing to suddenly fail at. A 6-year-old can sometimes hit a golf ball further than a golfer with 20 years of experience. Under pressure, an expert player can miss putts that a beginner would make.
This is not a flaw in the design. It’s a fundamental feature of how biological intelligence works. The brain solves control problems by exploiting structure, learning patterns, and operating automatically. These same features make it fragile to disruption—overthinking, excessive conscious attention, and fear can all disrupt automatic performance.
The lesson: the brain is not a universal, robust controller. It’s a specialized, evolved system optimized for learning and automatic execution. It’s fast and efficient because it’s automatic. It’s automatic because it has learned patterns from thousands of repetitions. Those same patterns can be disrupted.
Understanding this—that power and fragility are two sides of the same coin—is perhaps the most important insight motor control science offers (Nicholson 2005).
The brain solves the golf swing by exploiting structure: synergies (reducing from 200 muscles to 5 dimensions), feedforward control (pre-planning to avoid feedback delays), impedance control (commanding stiffness rather than position), and physics exploitation (letting drift do the work).
The brain learns three things through practice: the drift field (physics), the control effectiveness of muscles, and the optimal motor policy for achieving goals.
The brain is sample-efficient, building accurate models from 10,000–100,000 repetitions, while AI methods often require millions. The brain achieves this through built-in structure, hierarchical control, and rich sensory feedback.
Artificial intelligence could benefit from the brain’s approaches: hierarchical control, model-based planning, embodied exploitation of physics, and prediction-driven learning.
The golf swing is an ideal model system for studying intelligence because it’s complex, measurable, learnable, and has interesting failure modes.
The yips and choking reveal that motor control is fragile. The same automaticity that makes skilled performance powerful makes it vulnerable to disruption.
The ultimate irony: the most impressive control achievement your brain makes is a golf swing. And yet, you can blow it.
Chapter Exercises
Calculate the control problem dimensionality. If you naively controlled 200 muscles at 7 activation levels each, across 30 time points, how many possible swings would there be? Express in powers of 10.
Explain how synergies reduce the dimensionality problem. If synergies reduce from 200 muscles to 5 synergies, and you control each synergy’s activation at 7 levels across 30 time points, how many swings are possible? Compare to the naive case.
Why can the brain exploit the drift field to simplify control? How does understanding drift change the problem from a control perspective?
What are the three constraints that make feedback control impossible during the downswing? (Hint: latency, timing, and sensory delay.)
Explain impedance control. Why might commanding stiffness be simpler than commanding exact position?
How does the brain demonstrate that it learns physics? What experiments provide evidence?
Compare model-based and model-free reinforcement learning. Which approach is the brain using? Which approach is modern AI typically using?
What could AI systems learn from the brain’s approach to motor control? List three specific strategies.
Explain the yips using the setpoint hypothesis. Why might a corrupted reference trajectory lead to jerky, unstable movements?
According to the conscious processing hypothesis, why do expert golfers sometimes choke under pressure? How can this be prevented?
Design an experiment to test whether golfers can consciously control their putting stroke or whether conscious attention disrupts automatic performance.
The nervous system modulates output at 10–40 Hz while the shaft’s first bending mode is only 3–5 Hz, so the controller is faster than the plant. Explain why the golfer still cannot correct the swing in flight, in terms of loop delay against movement duration. What does the brain do instead of correcting?
What would a robot need to match human golf performance? List at least 5 capabilities and explain why each is difficult.
Explain the paradox: your brain solves an extraordinarily complex control problem while also being vulnerable to failure under pressure. Why are these two things related?