Chapter 14: The Secret to Dishwashing Robots

Core Idea

The motor cortex is usually described as the source of motor commands. Bennett, following Karl Friston, argues that it makes predictions, not commands. It models the body the way the aPFC models navigation and is wired to make its predictions come true. Its original job, in early placental mammals, was motor planning: simulating careful movements ahead of time and using those simulations to train the basal ganglia. The frontal neocortex and basal ganglia form a hierarchy of goals. Each level has a slow, flexible simulator in the neocortex and a fast, automatic version in the basal ganglia. This motor system is where the “secret to dishwashing robots” lies.

Frameworks Introduced

• The three problems with “motor cortex equals motor commands”:
  1. Motor cortex columns have the same microcircuit as the rest of the neocortex, which builds generative models. Any account has to explain how that circuit could produce commands.
  2. The first mammals had no motor cortex, and they moved fine. It appeared tens of millions of years later, only in the placental lineage (figure 14.2: the first placental mammal has a motor cortex, and its descendants are rodents, primates, carnivora and ungulata; marsupials such as possums, kangaroos and koalas have none).
  3. Paralysis after motor cortex damage is unique to primates. Rats and cats with a damaged motor cortex still walk, hunt and eat.
• Predictions, not commands (Friston):
  • How: the motor cortex observes body movements through the somatosensory cortex beside it, explains them, and predicts what the body will do next. The wiring sends those predictions to the spinal cord, so “the motor cortex is wired to make its predictions come true.”
  • Compared with the aPFC: the aPFC predicts navigational paths (the animal will turn left), and the motor cortex predicts body-part movements (it will place its left paw exactly on a platform).
  • Why it works: it reuses one generative circuit for a new job, with no new algorithm needed. This is the idea of embodiment, a full model of the body that can be simulated and adjusted over time.
• A hierarchy of goals: a balance of simulation and automation (figure 14.4):
  • How: the aPFC sets high-level goals (“drink water”) from amygdala and hypothalamus input. The premotor cortex turns them into subgoals, and the motor cortex turns those into sub-subgoals (“Position my index finger here and my thumb over here.”). Each frontal level has a loop through its own part of the basal ganglia and thalamus. In the figure, the frontal neocortex is labeled simulation and the basal ganglia automation.
  • Why it works: effort is spread out. The aPFC need not care about finger positions, and the motor cortex need not care why the cup is being lifted. Lower levels can respond flexibly to obstacles while serving one shared goal.

	High-level goals	Low-level goals
Simulation	Agranular prefrontal cortex: simulates navigational paths. Asks “do I want rigatoni or would I rather diet?” Damage impairs planning of navigational routes.	Motor cortex: simulates body movements. Asks “how do I configure my fingers to play this C chord that I just learned on guitar?” Damage impairs learning new motor skills and executing fine motor skills.
Automation	Front part of basal ganglia: automatically pursues a high-level goal in response to a stimulus. Produces habitual cravings. Damage can cure drug addiction.	Back part of basal ganglia: automatically executes a motor skill in response to a stimulus. Produces habitual motor responses. Damage impairs executing learned skills and forming motor habits.

Key Concepts

• Motor cortex: a thin band at the edge of the frontal cortex that maps the whole body. It gives most space to the body parts under skilled control (in primates, mouth and hands). Figure 14.1 shows the trunk, arm, hand, face and tongue areas running down the band, mirrored by the somatosensory cortex beside it.
• Motor planning: the motor cortex’s original role. Nonprimate mammals with motor cortex damage struggle with skilled movements, such as placing a paw on a thin branch, reaching through a small hole, stepping over an obstacle that is out of sight, or placing a foot on an uneven platform. They also cannot learn new movement sequences, though they can still perform sequences learned before the damage.
• Premotor cortex: the middle level that turns goals into subgoals. Damage can cause alien limb syndrome (limbs seem to move on their own) and utilization behavior or field-dependent behavior (drinking from empty cups, putting on other people’s jackets) because low-level goals are no longer constrained from above.
• Cravings: produced by the front part of the basal ganglia (“You come home and smell rigatoni”). The aPFC is what makes you pause and ask whether you want to pursue them.
• Mental rehearsal: imagining a movement activates the motor cortex. It substantially improves speaking, golf swings and surgical maneuvers. In one experiment, a seated person’s posture got worse when she heard sentences like “I get up, put on my slippers, go to the bathroom,” but not other sentences.
• Existence proof: as bird wings showed that flight was possible, mammal motor skills show what robots could do.

Mental Models

• Think of the motor cortex as the aPFC for the body. It uses the same predict-then-fulfill trick for paws and fingers instead of routes.
• Simulate while learning, and automate once learned. The motor cortex is needed to learn a new movement but not to perform a well-learned one, because the basal ganglia has taken over.
• Think of every goal level as having two systems. The neocortex is slower but more flexible, for training, and the basal ganglia is faster but less flexible, for well-trained paths and movements.
• Use damage location to predict behavior. Damage at high levels (aPFC or front basal ganglia) makes rats less sensitive to goals, so they keep pressing for food they no longer want. Damage at low levels makes them more goal-sensitive and unable to form motor habits.

Anti-patterns

• Treating the motor cortex as the controller of movement. That is true only in primates. In the first mammals it did not exist, and in rats and cats it is for planning.
• Building robots that need massive computing for every movement: the human brain runs on about the same energy as a lightbulb.
• Assuming reptiles plan movements: lizards running over obstacles do not anticipate them or adjust their forelimbs, and few reptiles live in trees.

Rules of Thumb (author’s wording)

• “the motor cortex was originally not the locus of motor commands but of motor planning.”
• “Once a movement is well learned, the motor cortex is no longer needed.”
• “The secret to dishwashing robots lives somewhere in the motor cortex and the broader motor system of mammals.”

Worked Example

Lesioning the front of the basal ganglia in heroin addicts (a study in China).

1. Prediction from the hierarchy: the front part of the basal ganglia automatically links cues to high-level goals, which is what produces cravings. Addicts show extreme activation there when they see drug cues.
2. The relapse rate for heroin addicts is very high, by some estimates as high as 90 percent.
3. The study lesioned the front part of the basal ganglia in the most severe addicts.
4. Relapse dropped to 42 percent. Patients lost the automatic pursuit of cues and the cravings it produces, though the surgery has many side effects and is highly controversial and of questionable ethics.
5. Reading: automated high-level goals live in the front basal ganglia, while resisting them takes extra aPFC activity that simulates the costs.

A second example: a cat stepping across posts (figure 14.3). A cat places each paw on a row of posts of different heights. This is the kind of planned movement that cats struggle with after motor cortex damage. Recordings show the motor and somatosensory cortex activate before such a precision movement, even when the obstacle is only known and not seen, and stay active until it is done.

Key Takeaways

1. The motor cortex predicts movements and is wired to make its predictions come true, which is the same algorithm as the rest of the neocortex.
2. It appeared in placental mammals for planning fine movements, and only primates came to need it for movement in general.
3. Frontal regions form a hierarchy of goals: aPFC, then premotor, then motor cortex.
4. Each level pairs neocortical simulation with basal ganglia automation. New skills start slow and deliberate and then become automatic.
5. Robots built this way would learn skills on their own, adjust in real time, work out subgoals from high-level goals, and learn faster over time by reusing low-level skills.

Connects To

• Ch 6: the basal ganglia’s actor-critic learning, which the frontal cortex now trains.
• Ch 11: the generative microcircuit the motor cortex reuses.
• Ch 13: active inference in the aPFC, habits, and the missing layer 4.
• Ch 15–18: primates, where the motor cortex becomes required for movement.

Summary of Breakthrough #3: Simulating

The main new brain structure of early mammals was the neocortex, and with it came simulation.

• The sensory neocortex evolved and created a simulation of the external world (a world model).
• The aPFC evolved as the first frontal neocortex. It simulated the animal’s own movements and internal states (a self model) and constructed “intent” to explain its own behavior.
• Together, the aPFC and sensory neocortex let early mammals pause and simulate things not currently happening. This is model-based reinforcement learning.
• The aPFC somehow solved the search problem by choosing which paths to simulate and when.
• Simulation enabled vicarious trial and error: choosing a path based on imagined outcomes.
• Simulation enabled counterfactual learning, a more advanced answer to the credit assignment problem based on causal relationships.
• Simulation enabled episodic memory: recalling past events and using them to adjust behavior.
• In later mammals the motor cortex evolved, which let them plan and simulate specific body movements.

Our ancestors “weaponized the imaginarium to survive,” outplanning dinosaurs. Like a modern cat, they could look at branches and plan where to put their paws. Other vertebrates of the time, like modern lizards and fish, could move fast, remember patterns, track time and learn through model-free reinforcement learning, “but their movements were not planned.” The next breakthrough, Bennett says, has been the hardest to reverse-engineer in AI, and it is a feat we rarely call intelligence.