1. Structured Intelligence or End-to-End Intelligence?
1.1 From The Society of Mind to Deep Learning
Marvin Minsky's The Society of Mind proposed that intelligence might emerge from the interaction of many smaller cognitive processes.
Historically, this idea was associated with symbolic and modular AI. Intelligence was decomposed into functions, rules, and subsystems. Human designers decided which module performed which task and how information moved among them.
Deep learning moved almost in the opposite direction.
Instead of specifying the internal structure of intelligence, we learned representations directly from data.
GPT pushed this philosophy remarkably far. Language understanding, factual knowledge, abstraction, reasoning, planning, style, and tool use can coexist inside one large learned model.
It is therefore tempting to conclude:
perhaps intelligence does not need architecture.
Perhaps scale itself is the architecture.
But this conclusion mixes two different claims.
The first is:
The internal representation of intelligence should be learned rather than manually specified.
The second is:
All intelligent functions should reside inside one undifferentiated controller.
The success of the first does not prove the second.
A system can be end-to-end learned while still being structurally organized.
Its components do not need to be hand-coded symbolic modules. Each component can itself be a powerful learned model.
So the real alternative is not:
It may instead be:
1.2 World Intelligence and Body Intelligence
For embodied intelligence, one particularly useful distinction may be between two broad domains.
The first deals primarily with the external world:
objects,
places,
other animals,
prey,
predators,
other people,
goals,
future possibilities,
and plans.
It asks:
What is happening in the world?
What may happen next?
What should I do about it?
We can call this:
The second deals primarily with the body:
posture,
balance,
joint state,
contact,
force,
locomotion,
breathing,
internal regulation,
and rapid physical execution.
It asks:
What is happening to my body?
How do I keep it stable and functional?
How do I execute the requested behavior?
We can call this:
These are not strict anatomical categories.
By “World Intelligence,” we roughly mean the cortical/conscious side of intelligence: perception, explicit reasoning, planning, prediction, and goal selection.
By “Body Intelligence,” we mean the broader subconscious body-control system: brainstem, spinal cord, cerebellar, autonomic, local sensorimotor, and peripheral regulatory circuits.
The crucial point is:
Body Intelligence is not a passive actuator layer.
It senses.
It maintains state.
It adapts.
It regulates.
It stabilizes.
It solves difficult control problems of its own.
The body is not merely controlled by intelligence.
The body contains intelligence.
2. The Engineering Problem
Why does this architectural distinction matter now?
Because robotics exposes limitations that are easy to ignore in purely digital AI.
2.1 Vision Is Computationally Expensive
A state-based locomotion policy can be surprisingly small.
Stable walking, running, or disturbance recovery can often be generated by relatively compact neural policies using structured inputs such as:
These variables are already closely related to the physical control problem.
Vision is very different.
A visual system begins with pixels and must infer:
objects,
geometry,
depth,
motion,
semantics,
context,
and sometimes intent.
Modern visual foundation models may therefore contain hundreds of millions of parameters.
The asymmetry is not merely a temporary engineering accident.
It reflects different information structures.
Motor sensing is already highly task-relevant.
Vision has to discover relevance from a massive spatial input.
This naturally suggests:
2.2 The Real-Time Problem
If the entire VLM or VLA must run at motor-control frequency, real-time execution becomes difficult.
The natural engineering response is often:
use a faster GPU.
Build a larger accelerator.
Increase onboard compute.
But this may answer the wrong question.
The deeper question is:
Why must the VLM appear in every motor-control cycle at all?
A more feasible architecture may be asynchronous and multi-rate.
For example:
while:
The motor policy does not wait for a new visual inference on every step.
It continues operating from proprioception, force, contact, and local state.
The visual system periodically updates:
target,
direction,
desired speed,
task mode,
trajectory,
or constraints.
Thus:
A useful principle is:
Vision should update intent, not clock the body.
2.3 The Vision-Robustness Paradox
Vision creates another engineering problem.
When perception is unreliable, we add more cameras.
Head cameras.
Wrist cameras.
End-effector cameras.
External cameras.
This reduces blind spots.
But more cameras also mean:
more data,
more synchronization,
more calibration,
more fusion,
more computation,
and potentially more latency.
So there is a paradox:
The attempt to make visual control more reliable may make the entire control system more computationally dependent on vision.
This raises the architectural question again:
should the fastest layer of physical control really depend on the heaviest perceptual subsystem?
2.4 Training May Also Be Architecturally Wrong
The second engineering problem is training.
Robot learning increasingly relies on image and video data.
The implicit learning pipeline becomes:
But video mainly provides:
task structure,
semantics,
approximate trajectory,
timing,
and planning information.
It does not directly provide many quantities needed for dynamic control:
contact force,
joint torque,
friction,
slip,
compliance,
actuator saturation,
disturbance recovery.
A useful distinction is:
Video teaches what motion looks like. Interaction teaches how to make it work.
A weak person does not learn to jump higher simply by watching more jumping videos.
A cyclist can learn racing lines from race-car videos but cannot thereby learn race-car dynamics.
The physical system being controlled matters.
This suggests a different curriculum:
First learn how to move.
Then learn how vision modifies movement.
Then learn how high-level intelligence selects and combines those skills.
2.5 Motor Skills Have Their Own Data
A motor system is not data-poor.
Biological systems have:
proprioception,
touch,
joint sensing,
muscle-related feedback,
balance,
pressure,
contact,
slip.
Robots have:
encoders,
IMUs,
force/torque sensors,
tactile sensors,
motor current,
joint velocity,
contact detection.
This sensing has limited knowledge of the distant environment.
But it contains rich information about the body and its dynamics.
Hence:
Vision is globally rich but dynamically indirect.
Motor sensing is environmentally narrow but dynamically rich.
For narrow motor tasks, fast sensing produces dense feedback.
Dense feedback creates many opportunities for correction.
Thus:
Fast sensing enables fast learning.
For some motor skills, feedback bandwidth may be as important as offline dataset scale.
3. Biological Perspective: Why Intelligence May Be Asynchronous
The engineering argument leads to a deeper question.
Is asynchronous control merely a workaround for current hardware?
Or is it a more general consequence of physical and biological constraints?
3.1 Biology Obeys the Same Physics
Biological systems face the same fundamental physical constraints as engineered systems:
finite energy,
finite bandwidth,
signal propagation delay,
sensor noise,
limited computational resources,
and actuator dynamics.
Neural communication is not free.
Long-distance signaling takes time.
Computation consumes metabolic energy.
Therefore, if routing every fast motor update through the most computationally expensive perceptual machinery is inefficient for a robot, biology has no exemption from the same physical trade-offs.
This suggests:
Asynchronous control may not be an engineering compromise. It may be a natural architecture under physical constraints.
3.2 A Developmental Clue: The Brain May Itself Be Composite
A striking recent result makes this architectural question even more interesting.
In September 2026, Jokhai, Dundes and colleagues reported in Nature Neuroscience that the developing mammalian brain may not arise from one common neural ectoderm progenitor.
Instead, lineage tracing in mouse embryos identified two parallel progenitor populations emerging during gastrulation.
The anterior neural ectoderm is committed to forming the:
while the posterior neural ectoderm is committed to forming the:
Experiments with human pluripotent stem cells supported the same distinction. The two progenitor populations showed different accessible chromatin landscapes and resisted signals attempting to convert one regional identity into the other.
The authors therefore proposed that:
the brain may be a composite organ derived from two lineage-restricted progenitors. [15]
They further suggested that this dual developmental organization may have deep evolutionary conservation.
This result should not be overinterpreted.
It does not prove that:
and:
The midbrain, for example, belongs to the anterior developmental lineage but has important sensory and motor functions. And Body Intelligence in our framework includes spinal, autonomic, cerebellar, peripheral, and other systems beyond the hindbrain.
But the result challenges an important implicit assumption:
Even the brain itself may not be developmentally one homogeneous organ that later differentiates internally.
Instead, major parts of the nervous system may begin as parallel lineages and later become tightly integrated.
That is architecturally suggestive.
Evolution does not necessarily seem to build intelligence by starting with one universal processor and assigning it every task.
It may combine neural systems with different developmental histories, capabilities, and responsibilities.
For AI, this leads to a provocative analogy:
Perhaps artificial intelligence also does not need to begin from one universal latent system that later learns every function. It may be more natural to develop distinct learned intelligences for fundamentally different problem domains and then learn how to connect them.
This is an analogy, not a biological proof of AI architecture.
But it strengthens the possibility that composite intelligence is not an engineering defect. It may be a very old design principle.
3.3 Vision Is Far-Field and Predictive
The latency of vision may make sense because vision and body sensing solve different spatial problems.
Vision is primarily a far-field, exteroceptive sense.
Consider a fish.
It may see prey or a predator before physical contact occurs.
Vision asks:
What is it?
Where is it?
Where is it moving?
Should I approach or escape?
Because the relevant object may be some distance away, the animal has a certain temporal margin.
Visual processing can therefore tolerate more latency.
The value of vision is precisely that it provides advance information.
Thus:
Vision is far-field and predictive.
3.4 Force and Proprioception Are Near-Field and Corrective
Force, touch, and proprioception are almost the opposite.
They tell the organism little about distant objects.
But they tell it in great detail:
Is the body tilting?
Has contact occurred?
Is the foot slipping?
Is a fin being loaded?
Is the joint moving correctly?
Is the force increasing?
When these signals become important, physical interaction is often already happening.
The time margin is much smaller.
Therefore:
Force sensing is near-field and corrective.
This suggests a functional division:
while:
3.5 Spatial Bandwidth and Temporal Bandwidth
This leads to another hypothesis.
Motor intelligence may computationally resemble auditory and haptic processing more than visual processing.
Not because they are the same biological system.
But because the structure of their information is similar.
Sound begins with mechanical pressure variation.
Touch begins with mechanical deformation, force, vibration, and contact.
Both can contain dense temporal information.
Vision instead begins with a huge spatial array and must derive structure from it.
So the contrast may be:
versus:
Raw force, vibration, and tactile sensing can plausibly operate at kilohertz-scale sampling rates in engineered systems, even if the learned motor policy itself updates more slowly.
The exact frequency is not the key point.
The information geometry is.
Vision is high-dimensional in space.
Motor sensing is high-resolution in time.
3.6 World Intelligence and Body Intelligence
This suggests a broader biological abstraction.
The cortex-like, conscious side of intelligence is mainly concerned with:
environment,
other agents,
prey,
predators,
goals,
future possibilities,
and planning.
This is:
The subconscious body-control system is mainly concerned with:
body state,
posture,
balance,
gait,
swimming,
contact,
muscle coordination,
breathing,
circulation,
and internal regulation.
This is:
The important point is that both are autonomous.
The lower system is not a dumb interface.
It contains complex sensing, feedback, state estimation, adaptation, and regulation.
The heart can generate rhythm locally while being modulated by broader autonomic control.
Respiration depends on interacting brainstem, sensory, mechanical, and autonomic loops.
Posture is continuously regulated through vestibular, proprioceptive, and contact feedback.
Locomotion requires coordination across large numbers of muscles and joints.
Swimming requires whole-body rhythmic dynamics interacting with fluid forces.
These are not simple commands.
They are sophisticated control problems.
3.7 Not Smart Brain and Dumb Body
The useful architecture is therefore not:
It is closer to:
World Intelligence may say:
Escape from the predator.
Body Intelligence solves:
How do I generate stable swimming under the current hydrodynamic conditions?
World Intelligence may say:
Walk toward that tree.
Body Intelligence solves:
How do I coordinate joints, maintain balance, change gait, and recover from disturbances?
The higher system owns much of the meaning.
The lower system owns much of the physics.
3.8 Biological Evidence from Local Motor Circuits
Locomotor research provides further clues.
Central pattern generators and spinal locomotor circuits show that substantial rhythmic structure can be generated locally under higher-level modulation.
Recent connectomics also reveals nervous systems composed of highly structured subsystems rather than one homogeneous controller.
A useful abstraction is:
The higher system decides:
whether to move,
where to move,
how fast to move.
The lower system solves much of:
how to produce and stabilize the motion itself.
4. Engineering Advantages of the Biological Design
If this interpretation is correct, biology suggests more than a metaphor.
It suggests an engineering architecture.
4.1 Multi-Rate, Asynchronous Intelligence
Instead of forcing every component onto one control clock, use multiple timescales.
For example:
The exact numbers are robot-dependent.
The principle is:
Each subsystem should operate at the bandwidth appropriate to its problem.
This reduces computation.
It reduces latency pressure.
It reduces dependence on fragile visual feedback.
It also makes system design more scalable.
4.2 Motor Primitives as Autonomous Agents
A motor primitive should not be imagined as a prerecorded trajectory.
It can itself be an adaptive intelligent system.
A running primitive can learn across:
different speeds,
slopes,
surfaces,
loads,
disturbances.
Its objective may be:
Given a desired direction and speed, remain dynamically stable and recover from perturbation.
The higher-level system does not specify every joint.
It says:
faster,
slower,
turn left,
stop,
change terrain mode.
Thus:
This is not command in the traditional software sense.
It is:
adaptive delegation.
4.3 Hierarchy as Information Compression
Suppose a robot contains actuators.
A centralized controller may need to output:
But if a lower-level learned system maps a compact latent condition into coordinated behavior,
then the higher-level system only needs:
The high-level signal may consist of:
speed,
direction,
goal,
mode,
urgency.
Thus hierarchy is not merely organizational.
It is a mechanism for information compression.
The more capable the lower-level agent becomes, the less information the higher level needs to transmit.
4.4 Skill as a Learned Manifold
The space of all possible motor outputs is enormous.
But valid behavior occupies only a structured subset.
Stable walking occupies a small subset of all possible joint trajectories.
Useful grasping occupies a subset of all possible hand configurations.
Music occupies a structured subset of all possible waveforms.
A skill can therefore be viewed as a learned manifold:
The lower-level agent learns this manifold.
Its latent state:
acts like a set of intrinsic coordinates.
The higher-level agent does not need to understand those coordinates directly.
It only needs a useful way to access the manifold.
4.5 Metric as Indexing Geometry
This leads to the idea of a metric.
Two gait states may be very different internally.
But if both produce stable forward motion at , they may be close from the perspective of the higher-level system.
Conversely, two internally similar states may be far apart if one succeeds on a slope and the other fails.
Distance is therefore task-dependent.
The metric organizes the lower-level skill space according to variables meaningful to the upper level.
For example:
speed,
direction,
distance,
terrain,
target.
Thus:
Pretraining learns the manifold.
Post-training shapes its metric.
The metric organizes and indexes the manifold.
The interface aligns one manifold with another.
The prompt selects a region or direction.
The agent executes the trajectory.
4.6 Learned Alignment Between Agents
Different agents need not share the same latent representation.
Suppose:
and:
There is no need to require:
Instead we can learn a bridge:
This bridge may be:
an adapter,
a projection head,
a small transformer,
a latent translator.
Its role is representation alignment.
The internal states remain different.
The interface learns how one intelligence should communicate with another.
This is important because the receiving system is itself intelligent.
It does not need a perfectly explicit instruction.
A sufficiently capable receiving agent can reconstruct detailed execution from a compressed high-level condition.
5. Current AI Trends and Their Relation to This Idea
This architecture does not reject modern VLMs, VLAs, or world models.
It changes the role we assign to them.
5.1 VLMs and VLAs
VLMs are extremely useful for:
scene understanding,
semantic reasoning,
object recognition,
task interpretation,
goal selection,
visual planning.
They are natural components of World Intelligence.
The architectural question is not:
Are VLMs useful?
Clearly they are.
The question is:
Should a large VLM directly own the fastest motor-control loop?
Our argument suggests that it often should not.
A more natural architecture is:
5.2 World Models
Recent embodied AI research is increasingly emphasizing world models and world-action models.
This is an important shift.
A VLM primarily helps answer:
What is happening?
A world model tries to answer:
What will happen if I do this?
It therefore extends:
physical reasoning,
prediction,
counterfactual simulation,
long-horizon planning,
understanding of environment dynamics.
From our perspective, a world model is a more powerful form of World Intelligence.
It provides a better model of the environment.
It can imagine future states.
It can compare alternative actions.
It can plan farther ahead.
But this does not automatically solve the body-control problem.
Indeed, deeper prediction and longer rollout may increase computation and reasoning time.
So the architecture may remain:
The world model may predict that the floor is slippery.
But once the foot actually slips, the local motor system still has to respond immediately.
Thus:
A better brain does not eliminate the need for a spinal cord.
The real question is not:
VLM or World Model?
It is:
How should either cooperate with autonomous Body Intelligence?
5.3 Current Robotics Already Contains Pieces of This Idea
Many existing research directions already point toward parts of this architecture:
hierarchical reinforcement learning,
options and temporal abstraction,
central pattern generators,
dynamical movement primitives,
latent skill spaces,
proprioceptive locomotion,
rapid motor adaptation,
vision-conditioned locomotion,
language-conditioned motor primitives,
and learned skill embeddings.
The claim is therefore not that no one has explored hierarchy.
Rather:
Many apparently separate techniques may be manifestations of one deeper architectural principle.
The deeper principle is that intelligence may naturally organize around:
different sensing modalities,
different timescales,
different latent spaces,
different control responsibilities,
and learned interfaces between them.
The recent discovery of parallel developmental brain progenitors adds an intriguing biological parallel: even the nervous system itself may be assembled from major neural lineages that are distinct from very early development and later integrated into one functioning organism. [15]
5.4 Society of Agents
This brings us back to Minsky.
The members of a modern Society of Agents need not be small symbolic rules.
Each member can itself be a powerful learned intelligence.
A Vision Agent understands the visible environment.
A World Model Agent predicts possible futures.
A Language Agent handles semantic reasoning.
A Locomotion Agent maintains stable movement.
A Manipulation Agent learns contact-rich dynamics.
A Tactile Agent performs high-bandwidth correction.
An Autonomic Agent regulates internal bodily state.
Each can possess its own:
representation,
sensor modality,
training data,
timescale,
and computational architecture.
They need not share one internal language.
They need only learn how to cooperate.
Conclusion
Deep learning taught us that intelligence should not be manually decomposed into rigid symbolic rules.
But it does not follow that intelligence should be architecturally monolithic.
Embodied intelligence faces multiple fundamentally different problems.
Vision is:
far-field,
spatially rich,
predictive.
World models are:
future-oriented,
planning-intensive,
computationally heavy.
Motor sensing is:
near-field,
mechanical,
temporally dense,
corrective.
Body control is:
fast,
local,
closed-loop.
Biology offers an additional clue. Recent developmental evidence suggests that even what we call “the brain” may be a composite organ arising from two parallel lineage-restricted progenitor populations rather than a single uniform progenitor. [15]
That does not prove that artificial intelligence should have exactly two systems.
But it weakens the intuition that a unified intelligent organism must arise from one homogeneous computational architecture.
The future of embodied AI may therefore not be:
It may instead be:
World Intelligence asks:
What is happening out there, what may happen next, and what should I do?
Body Intelligence asks:
What is happening to me, and how do I keep the body stable while executing the task?
The two communicate through compact learned interfaces.
The higher layer specifies intent.
The lower layer owns much of the execution.
This is not continuous micromanagement.
It is:
adaptive delegation.
The central question may therefore no longer be:
How large should intelligence become?
But:
What kind of intelligence belongs at each spatial and temporal scale, and how should those intelligences cooperate?
Minsky called it a Society of Mind.
For modern embodied AI, perhaps a more appropriate name is:
The Society of Agents
A society in which every member is itself intelligent, each agent operates at the spatial, temporal, and informational scale appropriate to its own problem, and coordinated intelligence emerges through learned alignment, compact conditions, prediction, and adaptive delegation.
Selected References
- Minsky, M. The Society of Mind. 1986.
- Sutton, R. S., Precup, D., & Singh, S. “Between MDPs and Semi-MDPs: A Framework for Temporal Abstraction in Reinforcement Learning.” Artificial Intelligence, 1999.
- Ijspeert, A. J. et al. “Dynamical Movement Primitives: Learning Attractor Models for Motor Behaviors.” Neural Computation, 2013.
- Kiehn, O. “Decoding the Organization of Spinal Circuits That Control Locomotion.” Nature Reviews Neuroscience, 2016.
- Guertin, P. A. “Central Pattern Generator for Locomotion.” Frontiers in Neurology, 2013.
- Minassian, K. et al. “The Human Central Pattern Generator for Locomotion: Does It Exist and Contribute to Walking?” The Neuroscientist, 2017.
- Jain, D., Caluwaerts, K., & Iscen, A. “From Pixels to Legs: Hierarchical Learning of Quadruped Locomotion.” CoRL, 2021.
- Kumar, A. et al. “RMA: Rapid Motor Adaptation for Legged Robots.” RSS, 2021.
- Miki, T. et al. “Learning Robust Perceptive Locomotion for Quadrupedal Robots in the Wild.” Science Robotics, 2022.
- Tang, C. et al. “SayTap: Language to Quadrupedal Locomotion.” CoRL, 2023.
- Assran, M. et al. “V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning.” 2025.
- NVIDIA. “Cosmos World Foundation Model Platform for Physical AI.” 2025.
- Reuss, M. “Pretrained to Imagine, Fine-Tuned to Act: The Rise of World-Action Models.” NVIDIA Seattle Robotics Lab, 2026.
- Lin, A. et al. “Network Statistics of the Whole-Brain Connectome of Drosophila.” Nature, 2024.
- Jokhai, R. T., Dundes, C. E., et al. “Two Parallel Neural Ectoderm Progenitors Contribute to the Developing Brain.” Nature Neuroscience, published September 18, 2026.