Phenomenology, process and enaction

Merleau-Ponty and the body-subject

An agent is bodily oriented when its changing morphology continuously reshapes what is reachable, passable and actionable, so the world arrives as a field of “I can” before it becomes an explicit map.

TLDR

The lived body is more than an object inside a neutral world; it is the practical orientation from which a world becomes available.
The decisive mechanism is body-relative calibration: environmental geometry is converted into probabilities of successful action.
A physical shell, a camera and a body model do not suffice if their action boundaries stay stale when morphology or control changes.
An embodiment-substitution study tests orientation by measuring error, recovery, after-effects and transfer after an unannounced body change.
Passing the test supports adaptive bodily control, not claims of ownership, subjectivity, experience or consciousness.

The doorway did not move

A participant walks down a corridor and passes through a 72-centimetre opening without slowing. On the next trial, a lightweight frame widens the participant’s shoulders by 14 centimetres. The corridor, lighting and doorway are unchanged. What should happen?

The obvious answer is that the participant sees the same doorway, consults a new body measurement and chooses whether to turn. That answer hides a sequence: first construct a neutral picture of the world, then add the body, then calculate an action. It sounds sensible because engineers often build systems this way.

Yet the first few steps after the frame is attached expose something less orderly. The old gait can persist. The shoulders begin to turn too late. A scrape or near miss produces a correction. After several attempts, openings of the same width solicit a different approach. Nothing in the visible geometry changed. The practical meaning of the geometry changed because the acting body changed.

Maurice Merleau-Ponty’s term body-subject names the deeper structure visible in this small disturbance. The body is not only what a person can look at, measure or describe. It is also the tacit orientation through which near and far, possible and impossible, invitation and obstacle acquire their lived sense. His claim belongs to phenomenology, not to a laboratory model. But it creates a sharp engineering question: what would have to change inside a system for the same environment to become a different field of possible action?

This paper develops one answer. The causal work of embodiment lies in a closed calibration loop that maps morphology, current capability and sensorimotor feedback into action boundaries before explicit planning. The claim is narrower than saying perception contains no representations. It is stronger than saying a robot merely needs a body. It predicts a characteristic signature when the body is altered without warning: transient orientation error, feedback-sensitive recovery, selective transfer and, often, an after-effect when the old body returns.

The body is not consulted afterwards

Stand beside a table and notice the cup near its edge. You can report its colour, estimate its distance and name its shape. But before any of those judgements, the cup is already graspable from here, awkward from there, and unreachable without leaning from somewhere else. Its position is not only a coordinate. It is organised around a body with a certain reach, posture, balance and history of skill.

Merleau-Ponty’s target was a shared assumption in empiricist and intellectualist pictures of perception. One picture builds experience from sensory pieces. The other lets a judging mind organise those pieces. Both can leave the body as an object whose contribution is added after a world has already been constituted. In Phenomenology of Perception, he instead treats bodily motility as an operative way of being directed towards a situation. The body’s practical directedness is not a verbal intention such as “I shall now reach”. It is the background organisation that lets the cup solicit a reach at all.1 2

A spatial situation is more than a position

Imagine two descriptions of the same kitchen. The first is a surveyor’s plan. It places the cup at coordinate (2.1, 0.8), the counter 92 centimetres above the floor and the doorway 1.4 metres behind the observer. The second description begins from a seated person holding a hot pan. The cup is within the free hand’s reach, the counter is too high for a safe transfer, and the doorway lies behind a turn that the occupied arm makes awkward. The coordinate map is unchanged, but the practical world is not.

Merleau-Ponty’s point is not that metric space is false. It is that metric description abstracts from a more primitive spatiality of situation. Up and down, near and far, open and blocked are first organised around posture, orientation and possible movement. A body supplies a practical origin without needing to appear as a labelled origin on an inner map. When a person automatically ducks beneath a beam or shifts a load before entering a narrow passage, the adjustment can precede any reportable estimate of height or clearance.

This gives the paper a precise causal demand. A system counts as bodily oriented only if changing its current powers reorganises what the same scene makes possible. A database may contain perfect dimensions for the body and kitchen yet remain causally inert. A planner may query those dimensions only after a goal has been symbolically specified. Orientation is present when bodily capacity already shapes the candidate actions that reach deliberation. The substitution study therefore looks upstream of explicit choice, at the changing field from which choices are drawn.

The world as a body-relative field. The same object can remain metrically fixed while moving between immediate, effortful and unavailable action regions. The figure depicts the paper’s core mechanism, not a neural map. Illustrative.

Four bodies that should not be collapsed

Discussion becomes confused because “the body” can refer to at least four different things. A body-object has measurable dimensions and forces. A body image includes perceptual, conceptual and affective attitudes towards one’s body. In much contemporary cognitive science, a body schema denotes sensorimotor organisations used to guide action, although the term covers several models and its boundaries remain disputed.10 The body-subject is broader. It names the first-person bodily orientation through which a situation has practical sense.

Habit is a changed availability, not only a stored rule

Consider learning to type without looking at the keyboard. A novice remembers letters, searches for keys and issues one finger command at a time. With practice, a word can become a single directed movement whose details recede from attention. The keyboard has not moved and the fingers have not acquired new joints. What changed is the immediate organisation of possible action. Some sequences now arrive as available wholes, while an unfamiliar shortcut still interrupts the flow.

Merleau-Ponty treats habit as a bodily acquisition because skill alters the range of what can be done directly. This does not require a mystical memory in the muscles. Neural plasticity, predictive control and learned policies may all contribute. The phenomenological point concerns how the acquisition is lived: the skilled person does not normally calculate a route through joint space and then execute it. The route has become one of the body’s practical powers. Replace the keyboard layout without warning and this history becomes visible through systematic errors, rapid correction and after-effects when the old layout returns.

The same logic applies to a mobility aid, a vehicle or a remote manipulator. Mere possession does not integrate the tool. Integration is shown when distances, clearances and reachable futures are recalibrated around the coupled body-tool system. That criterion is demanding enough to test, yet modest enough to avoid claiming that every learned controller instantiates a lived body.

Body-object

Third-person properties that can be measured independently of the agent’s report.

shoulder width
joint limits
actuator torque
mass and inertia

Body-subject

The whole practical orientation in which “near”, “reachable”, “blocked” and “requiring care” already matter.

Not an extra internal object. It is the body as lived point of departure.

Body image and body schema

Related but separable constructs used to explain bodily awareness and action.

beliefs and appearance
felt ownership
postural state
action-oriented calibration

One body, four explanatory roles. Neural or computational body schemas may help implement bodily orientation, but no single schema automatically exhausts the phenomenological body-subject. Practitioner synthesis.

What disappears when proprioception disappears

Most people can close their eyes and bring a finger towards the nose because limb position is available without visual inspection. Proprioceptive signals contribute to position, movement and force sense, and to online correction.12 Severe deafferentation makes the normally tacit contribution visible. The well-documented subject Ian Waterman lost major touch and proprioceptive inputs below the neck while retaining motor output. He learned to move by visually monitoring and deliberately controlling his body. Movement remained possible, but it demanded sustained attention and compensation.11

The case does not show that an intact body schema is unconscious in one neat module and a body image consciously replaces it. The empirical taxonomy is messier. It does show why the body-subject cannot be reduced to a list of dimensions known explicitly. When a normally available bodily orientation is disrupted, an intelligent person can reconstruct action through effortful strategies. Explicit knowledge can compensate for orientation without becoming equivalent to it.

Thought experiment 1

The widened body

Keep the corridor, vision and goal fixed. Change only shoulder width. Then hide the modification under clothing so the participant cannot solve the task from an obvious visual cue. Does the first attempt use the old turning threshold? Does outcome feedback move the threshold? Does the update transfer from walking through an aperture to squeezing sideways? A genuine orientation mechanism predicts structured recalibration, rather than only a verbal acknowledgement that the body is wider.

Now separate three possibilities. In the first, the participant is told the new width and can repeat it accurately, yet the early turns remain timed for the old body. Explicit body knowledge has changed while practical orientation has not. In the second, the participant turns earlier on the first trial because the frame is visible or the instruction reveals the manipulation. Performance is correct, but the result does not show calibration from action. In the third, the first attempts expose the old boundary, relevant contact or near-contact drives recovery, and removal of the frame produces a temporary over-turn. Only the third pattern carries the trajectory expected from a recalibrated body-world relation.

The test should also include apertures that are obviously passable or impossible. If behaviour changes equally at every width, the participant may simply be acting cautiously because the apparatus feels unusual. Orientation predicts its largest change near the action boundary, where a small body change alters the probability of success. The manipulation therefore turns a philosophical distinction into a shape of data.

The argument in one pass

Chart · flow

flowchart TD
  A["The doorway did not move"] --> B["The body is not consulted<br/>afterwards"]
  B --> C["Orientation is a calibration<br/>problem"]
  C --> D["The substitution test"]
  D -. "re-test" .-> A
  classDef first fill:#0a1424,stroke:#D95545,stroke-width:2px,color:#e6f3ff
  classDef step fill:#0a1424,stroke:#3f7da0,stroke-width:1px,color:#e6f3ff
  class A first
  class B,C,D step

A structural route through Merleau-Ponty and the body-subject: from The doorway did not move to Continue through the canon. Reading use: follow the relation, then return to the surrounding argument for evidence and limits.

Orientation is a calibration problem

A doorway does not afford passage in isolation. Passage belongs to a relation among opening width, shoulder profile, compressibility, gait, load, speed, confidence and the cost of collision. James Gibson’s ecological psychology made such relations central through the concept of affordance.3 Classic experiments showed that stair-climbing and aperture decisions can be organised by body-scaled information rather than by absolute centimetres alone.4 5

Body scaling is only a start. Two people with equal shoulder width can differ in how much they rotate, compress, hesitate or recover from contact. Even one person varies across trials. A useful action boundary is therefore not always a hard line. Franchak and Adolph propose treating an affordance as a probability-of-success function, which separates the point of transition from variability around it.7

P(success | r) = 1 / (1 + exp[-k(r - θ)]) r is an environment-to-body ratio, θ is the calibrated transition point, and k controls how sharply success changes around that point. The equation assumes other conditions are held stable. Remove that assumption and θ must vary with posture, fatigue, load, surface and risk policy.

This equation is not a theory of lived experience. It is a compact way to expose the mechanism. The body changes the denominator. Practice changes the calibrated threshold. Uncertainty changes the slope. Decision cost changes where the agent chooses to act. The body does more than add data to a completed perception; it helps determine which environmental differences are action-relevant.

The body substitution is the causal intervention. Holding the external scene fixed while changing morphology isolates whether the action boundary is actually coupled to the body. Proposed experiment.

Calibration turns a body into an orientation

Calibration is the relation between information and the action system that uses it. A driver can see optical expansion yet brake too late if expansion is miscalibrated to current vehicle dynamics. A person can see a doorway accurately yet turn too late after a sudden change in body width. Fajen argues that successful visually guided action requires learning information about boundaries between possible and impossible actions, rather than only detecting optical variables.6

Recalibration is selective. Franchak found that exploration sufficient for one fitting task did not automatically recalibrate another. Participants needed action practice or feedback relevant to squeezing through doorways.8 Later work showed that the rate of recalibration depends on feedback conditions.9 This specificity matters. If any motor activity updated every action boundary equally, “orientation” would be another name for generic learning. Instead, the system must discover which bodily and environmental variables jointly govern the task.

Thought experiment 2

The telescoping arm

Give an agent an arm that can suddenly extend twice as far, but do not alter its camera image or symbolic map. Place an object just outside old reach and inside new reach. A planner told the new length can calculate the answer immediately. A sensorimotor orienting system must explore and recalibrate. Now retract the arm without warning. The planner changes when its parameter changes; the orienting system shows a transient after-effect. The difference lies in how capability becomes available, not in whether the final reach succeeds.

Add a second object at the old reach boundary and a third that is reachable only if the extended arm approaches from the side. A single length parameter predicts the first two but may fail on the third because controllability, joint limits and collision geometry also changed. If exploration updates every reach judgement by the same amount, the system has learned a coarse scale factor. If transfer follows the shared trajectory constraints, it has learned a richer action relation. The thought experiment therefore distinguishes possessing a dimension from reorganising a field of reachable possibilities.

Finally, give the planner correct morphology metadata but delay that metadata by ten trials. Give the orienting controller no metadata but immediate contact and endpoint feedback. Their recovery curves can cross. The planner is accurate once told; the controller may adapt before it can state what changed. That crossing is the operational meaning of bodily orientation preceding explicit representation.

Tools reveal plasticity without proving incorporation

Tool use offers a natural body substitution. In macaques trained to retrieve objects with a rake, Iriki and colleagues reported changes in bimodal neuronal receptive fields that extended along the tool or into the newly accessible space.13 Human tool-use studies have also reported changes in subsequent free-hand movement kinematics consistent with an updated arm representation.14

These findings support plastic action-oriented body coding. They do not show that the rake literally became phenomenological flesh, that ownership was transferred, or that one cortical field is the body-subject. The rubber-hand illusion makes the distinction even clearer. Synchronous vision and touch can shift reported ownership and felt location towards an artificial hand.15 Yet ownership reports and motor measures can dissociate. Body schema, body image, ownership and body-subject are related evidence streams, not interchangeable labels.

The signature is a trajectory, not a final score. Error at substitution, recovery with relevant feedback, task-specific transfer and washout after-effects reveal how an action boundary is calibrated. Values are schematic. Illustrative.

Thought experiment 3

The frozen film

Record the exact camera frames, proprioceptive readings and action outputs from a successful reach. Replay them through an identical processor, but disconnect its actions from the world. The replay produces the same data sequence and the same final command. Has it retained the same bodily orientation? A contingency-based account says no: the sequence is no longer governed by what the agent can do next.16 This thought experiment changes only counterfactual coupling. It shows why a log of sensor values is weaker evidence than an intervention on the live loop.

Move the target halfway through the replay. The live agent can detect the deviation, revise its trajectory and perhaps abstain. The replay continues towards the old location while still reproducing a sequence that was once successful. Actual inputs and outputs were sufficient to describe what happened in the recorded trial, but not what the system was able to do under a nearby alternative. Bodily orientation concerns this counterfactual readiness. It is a disposition maintained by a functioning loop, not a property that can be inferred from one frozen history.

This matters for machine evaluation because logged trajectories are easy to over-read. A model can generate plausible proprioceptive commentary from a trace, and a simulator can replay correct commands, without either controlling the body that produced them. Break the coupling, perturb the target or reverse the control law. The system that remains oriented should reorganise action according to the new consequences.

The strongest empirical objection

Some embodied-perception claims go further and say effort, intention or ability alters how spatial properties visually appear. That literature is contested. Firestone and Scholl argue that many proposed top-down effects can be explained by judgement, task demands, attention or low-level differences rather than a changed visual experience.17 The criticism should change the thesis, not be waved away.

The substitution test

How can we move from evocative examples to a discriminating experiment? Start by preserving the task and external world while substituting one causal property of the body or control loop. The intervention might widen a participant, extend a tool, alter a robot gripper, add actuator lag or remap a joystick. Then measure the full adaptation trajectory rather than final accuracy.

The proposed contribution is an embodiment-substitution study. Its unit of analysis is not the naked body and not the model alone. It is the configured loop containing morphology, sensors, action policy, feedback and the environmental task. The study compares systems that can reach the same endpoint through different mechanisms.

Four candidate mechanisms

A geometry-only classifier learns that apertures above 70 cm are passable. It can score well while body size stays fixed. A static body-scaled controller divides aperture width by a stored body width, which is better but fails if that stored value becomes stale. An adaptive orienting controller updates the body-relative boundary from prediction error. An explicit online re-planner receives or infers the new morphology and recomputes its model. The last two may show similar behaviour even though one is described as embodied calibration and the other as model-based estimation.

This is not a defect. It marks the experiment’s proper conclusion. The test can identify adaptive orientational function; it cannot prove that the implementation is non-representational. Further interventions are needed to separate candidate algorithms.

A metric constant can become an action discontinuity. The example converts one unchanged doorway into two different success distributions by altering only effective body width. Synthetic worked example.

orientation_gap(t) = meanₓ | P̂(success | x, body_estimateₜ) − P(success | x, body_trueₜ) | The mean is taken over a controlled set of environments x. A useful report includes the peak gap after substitution, trials to recover, residual gap, transfer to a new task and after-effect when the original body returns.

Test the configured loop, not the model in isolation. Bodily orientation depends on how world state, morphology, sensing, action and feedback constrain one another over time. Proposed system boundary.

A protocol that forces mechanisms apart

The baseline phase estimates each system’s action boundary and variability. The substitution phase changes one bodily property without changing the world. The exploration phase controls what feedback is available. The transfer phase changes the task while preserving the altered body. The washout phase restores the original body. A strong protocol also masks obvious visual identifiers, randomises trial order and includes a matched no-change condition.

The outputs are not just success and failure. Measure when the system initiates a turn, how much clearance it leaves, whether it abstains, how prediction confidence changes, and how quickly the internal or behavioural boundary moves. For a robot, log the body-state estimate, policy proposal, action, contact event and readback. For a human, combine action measures with confidence and carefully worded experience reports. Keep those evidence types separate.

Thought experiment 4

The body changed without changing shape

A remote robot keeps the same dimensions, camera and gripper, but 280 milliseconds of unpredictable control latency is introduced. The visible body is unchanged, yet reachable trajectories and safe stopping distances alter. If orientation concerns only morphology, nothing should change. If it concerns the live “I can”, the action field should contract or become more conservative. This intervention tests whether temporal controllability belongs inside the effective body.

Compare constant latency, randomly varying latency and a visual delay that leaves motor control untouched. A controller that merely notices degraded video may slow equally in all three conditions. A body-relative controller should change most where the delay alters the mapping between issued movement and resulting contact. It should widen stopping margins, lower approach speed near obstacles and recover when timing becomes predictable again. Restoring zero latency may then produce a brief under-reach or premature stop if the delayed control law had been internalised.

The rival explanation is generic risk aversion. To test it, include a non-bodily uncertainty condition with the same failure rate, such as intermittent target occlusion. If safety margins expand identically, the result says little about bodily orientation. If the pattern tracks controllability and transfers to tasks sharing the same temporal dynamics, effective embodiment includes the timing of action, not only the outline of the machine.

Correct behaviour can be generated by the wrong cause. Cue masking and independent body changes are needed to prevent a classifier from imitating calibrated orientation. Adversarial test design.

How the claim changes under pressure

Chart · state

stateDiagram-v2
  [*] --> Claim
  state "A spatial situation is more<br/>than a position" as Claim
  state "Four bodies that should not be<br/>collapsed" as Test
  state "Habit is a changed<br/>availability, not only a<br/>stored rule" as Revision
  Claim --> Test: expose mechanism
  Test --> Decision: evidence survives
  Test --> Revision: rival survives
  Revision --> Claim: narrow or rebuild
  Decision --> [*]

The revision loop begins with A spatial situation is more than a… and keeps counterexamples inside the reasoning path. Reading use: follow the relation, then return to the surrounding argument for evidence and limits.

Run the embodiment-substitution lab

The interactive lab below turns the argument into a synthetic experiment. It compares three controllers across baseline, an unannounced body change and washout. The stale controller never updates its body estimate. The adaptive controller updates from delayed, noisy feedback. The oracle re-planner receives the true body scale immediately. Each controller encounters the same generated apertures.

The chart reports orientation gap, defined as the absolute difference between predicted and true probability of success, averaged over controlled apertures. A positive result is not “the adaptive line is low”. The informative pattern is an error spike at substitution, recovery governed by feedback quality and adaptation rate, plus an after-effect when the old body returns. The oracle is a rival mechanism: it can perform well without a gradual sensorimotor trajectory.

Embodiment-substitution lab

Change morphology, feedback and loop timing. The simulation uses synthetic data and a seeded random generator.

Executable in this page

Stale controller

0.00

Peak post-change gap. No recovery is expected unless the world happens to make the stale rule harmless.

Adaptive controller

0 trials

Trials to halve the initial substitution gap. “Not reached” means feedback was insufficient.

Washout after-effect

0.00

Adaptive gap immediately after the original body returns.

Ready. Baseline trials 1–15, substituted body 16–55, original body restored 56–80.

An executable test of orientational adaptation. The lab exposes which assumptions create transient error, recovery and washout. It does not simulate consciousness or ownership. Synthetic simulation.

What the lab assumes

The environment is reduced to aperture ratios. True success follows a logistic curve. The adaptive controller has one latent estimate of body scale. Feedback quality, latency and noise determine how rapidly that estimate approaches the altered body. Real bodies violate every simplification: morphology is multidimensional, control is hierarchical, exploration is strategic, contact changes the world and goals alter acceptable risk.

Those simplifications are useful because they reveal the argument’s hidden commitments. If the body estimate is never used, substitution creates no structured gap. If the estimate is used but never updated, the gap persists. If the true scale is supplied as metadata, the oracle recovers without an error-driven trajectory. If delayed feedback weakens adaptation, the effective body includes temporal control, not just shape.

JavaScript · core update rule

function updateBodyEstimate(estimate, trueScale, options, randomNoise) {
  const latencyPenalty = Math.exp(-options.latencyMs / 300);
  const usableFeedback = options.feedbackQuality * latencyPenalty;
  const predictionError = trueScale - estimate;
  return estimate
    + options.adaptationRate * usableFeedback * predictionError
    + randomNoise * options.feedbackNoise;
}

Run the lab with high feedback and low latency. The adaptive controller should recover. Set feedback quality to zero and it should remain miscalibrated. Increase latency while keeping morphology fixed. Recovery slows because sensorimotor consequences arrive too late to be assigned cleanly. Make the body smaller rather than larger. The risk reverses: the stale controller becomes overly cautious instead of overconfident.

What a null result can and cannot mean

No recalibration after substitution does not have one interpretation. The altered property may be irrelevant to the chosen action. Feedback may be too weak, too delayed or too ambiguous to identify the change. The substitution may exceed the controller’s recoverable range, producing failure rather than learning. A cautious policy may already leave so much margin that the behavioural threshold barely moves. In a human study, novelty, discomfort or instruction may dominate the action relation. In a machine study, a frozen perception module may estimate the change correctly while an authority layer prevents policy updates.

A useful null therefore needs manipulation checks. Confirm that the body change altered true success near the baseline boundary. Confirm that the sensors could, in principle, detect relevant consequences. Include an oracle controller to show the task is solvable with the new body. Include a learning-positive control to show the adaptive mechanism can update under some condition. If all checks pass and the configured system still applies the old boundary, the result weakens the claim that bodily capacity shapes action online.

Positive adaptation is also insufficient on its own. A controller may infer the trial block from a sound, colour or operator routine. It may retrieve a pre-trained policy for the substitute body. It may receive the true morphology through hidden metadata. Randomise those cues, introduce novel intermediate changes and compare recovery under feedback ablation. The evidential target is not improvement after a change; it is improvement caused by the live body-world relation the thesis names.

A field protocol for humans or machines

Choose an action with a measurable boundary, such as passage, reach, grasp, braking or safe turning. Estimate a baseline success function across controlled environmental values. Substitute one bodily property without announcing the trial boundary. Make the change independent of visible labels. Provide different feedback conditions across groups or runs. Then test transfer and washout.

For humans, preregister whether the primary outcome is movement, judgement or experience report. For machines, version the controller, body configuration, sensor calibration, action policy and environment generator. Preserve trial-level receipts. In both cases, include a no-substitution control and an oracle condition. The oracle answers an important rival: perhaps the system needs accurate body data, not gradual sensorimotor calibration.

Observed pattern	Static geometry or cue rule	Adaptive orienting loop	Strongest conclusion permitted
Immediate post-change error	Possible if the learned rule is stale	Expected when substitution is unannounced	The old body-action relation no longer fits
Recovery depends on outcome feedback	Not expected unless the rule itself learns online	Expected, with rate tied to usable feedback	Action consequences contribute to recalibration
After-effect when original body returns	Weak unless state persists	Expected if the updated relation has inertia	A body-relative state was retained across trials
Transfer follows shared body relation	May follow visual or semantic similarity instead	Expected where tasks share control variables	Calibration is structured, not wholly generic
Performance survives cue masking	Shortcut is exposed	Expected if live consequences remain available	Behaviour is not explained by the masked cue
Ownership or first-person report	Not addressed	Not entailed	Requires a separate evidential protocol

The embodiment-substitution evidence matrix. Each observation narrows mechanisms but stops before phenomenological or consciousness claims. Decision instrument.

Machines can pass the functional test without becoming body-subjects

Robotics already demonstrates parts of this programme. Bongard, Zykov and Lipson built a robot that used sensorimotor evidence to construct and update self-models, then adapted behaviour after damage.18 Robotics reviews collect many related uses of body-schema language for self-calibration, multisensory integration and control.19 More recent work has used egocentric vision and action sequences to predict robot dynamics across morphologies and to recover after hardware anomalies.20

These are significant engineering achievements. They show that a system can learn relations between its body, action and changing sensory consequences. They do not show that the system lives its body as a subject. The philosophical comparison preserves one relation: both cases make action possibilities depend on an embodied point of control. It breaks at first-person givenness, existential significance and the experienced unity of a situation. Functional orientation is a researchable bridge; phenomenological identity is not licensed by the bridge.

Optional depth: pre-registration fields for a substitution study

Candidate boundary: list model, runtime state, sensors, effectors, morphology and environmental feedback included in the system.
Intervention: name the single altered property and verify that external task geometry is held constant.
Primary outcome: select movement threshold, prediction gap, collision rate, abstention, confidence or report before data collection.
Rival generators: implement cue lookup, static body scaling and oracle parameter update.
Transfer test: specify the relation expected to generalise and a visually similar task where it should not.
Null policy: state whether no adaptation rejects the mechanism, indicates insufficient feedback or shows the task did not depend on the altered property.

Change the body and watch the action field recalibrate

When evaluating an embodied human-machine system, do not ask only whether it has cameras, limbs, a world model or a self-description. Ask what happens when its body changes while the world stays fixed. Does the system continue applying a learned geometric rule? Does an operator manually update a parameter? Does a live loop detect prediction error and recalibrate the field of possible action? Does that calibration transfer according to shared body-environment relations?

This changes both design and language. In design, place body-state estimation, feedback timing and affordance calibration inside the tested system boundary. Measure recovery trajectories and after-effects, not only endpoint performance. Build rival generators that can imitate correct behaviour through visible cues or oracle metadata. In language, stop treating every robot with a chassis as evidence for the lived body.

Merleau-Ponty’s body-subject is not an antique synonym for a kinematic model. It is a reminder that a world is encountered from somewhere, with powers and limits that already organise significance. Engineering cannot simply inherit the phenomenological conclusion. It can, however, isolate one of its causal demands: a body must continuously matter to what the system can do next.

Choose differently: before granting an architecture the stronger label “embodied intelligence”, require an unannounced body-substitution test with feedback manipulation, transfer, washout and a matched shortcut model. Passing supports adaptive bodily orientation. Failing means the body is still peripheral to the intelligence being claimed.

Glossary

Body-subject: The lived bodily orientation through which a situation appears as practically meaningful, rather than the body considered only as an object.
Body schema: A family of action-oriented sensorimotor representations or processes. Definitions and boundaries vary across disciplines.
Affordance: A possibility for action defined by a relation among environmental properties, an agent’s capabilities and task conditions.
Calibration: The learned or maintained relation between available information and the action boundary it is used to control.
Orientation gap: The divergence between predicted and actual action success after a body or control-law change, evaluated across controlled environments.
Motor intentionality: Merleau-Ponty’s term for practical bodily directedness towards a situation before a reflective decision or explicit judgement.

References

Merleau-Ponty, M. (1945/2012). Phenomenology of Perception. Translated by Donald A. Landes. Routledge. Publisher record and DOI.
Toadvine, T. “Maurice Merleau-Ponty.” Stanford Encyclopedia of Philosophy. Authoritative overview.
Gibson, J. J. (1979/2014). The Ecological Approach to Visual Perception. Publisher record and DOI.
Warren, W. H. (1984). Perceiving affordances: Visual guidance of stair climbing. Journal of Experimental Psychology: Human Perception and Performance, 10(5), 683–703. DOI.
Warren, W. H., & Whang, S. (1987). Visual guidance of walking through apertures: Body-scaled information for affordances. Journal of Experimental Psychology: Human Perception and Performance, 13(3), 371–383. DOI.
Fajen, B. R. (2005). Perceiving possibilities for action: On the necessity of calibration and perceptual learning for the visual guidance of action. Perception, 34(6), 717–740. DOI.
Franchak, J. M., & Adolph, K. E. (2014). Affordances as probabilistic functions: Implications for development, perception, and decisions for action. Ecological Psychology, 26(1–2), 109–124. DOI.
Franchak, J. M. (2017). Exploratory behaviors and recalibration: What processes are shared between functionally similar affordances? Attention, Perception, & Psychophysics, 79, 1816–1829. DOI.
Franchak, J. M., & Somoano, F. A. (2018). Rate of recalibration to changing affordances for squeezing through doorways reveals the role of feedback. Experimental Brain Research, 236, 1699–1711. DOI.
de Vignemont, F. (2010). Body schema and body image: Pros and cons. Neuropsychologia, 48(3), 669–680. DOI.
Gallagher, S., & Cole, J. (1995). Body image and body schema in a deafferented subject. Journal of Mind and Behavior, 16(4), 369–390. Author-hosted preprint.
Proske, U., & Gandevia, S. C. (2012). The proprioceptive senses: Their roles in signalling body shape, body position and movement, and muscle force. Physiological Reviews, 92(4), 1651–1697. DOI.
Iriki, A., Tanaka, M., & Iwamura, Y. (1996). Coding of modified body schema during tool use by macaque postcentral neurones. NeuroReport, 7(14), 2325–2330. DOI.
Cardinali, L., Frassinetti, F., Brozzoli, C., Urquizar, C., Roy, A. C., & Farnè, A. (2009). Tool-use induces morphological updating of the body schema. Current Biology, 19(12), R478–R479. DOI.
Botvinick, M., & Cohen, J. (1998). Rubber hands ‘feel’ touch that eyes see. Nature, 391, 756. DOI.
O’Regan, J. K., & Noë, A. (2001). A sensorimotor account of vision and visual consciousness. Behavioral and Brain Sciences, 24(5), 939–973. DOI.
Firestone, C., & Scholl, B. J. (2016). Cognition does not affect perception: Evaluating the evidence for “top-down” effects. Behavioral and Brain Sciences, 39, e229. DOI.
Bongard, J., Zykov, V., & Lipson, H. (2006). Resilient machines through continuous self-modeling. Science, 314(5802), 1118–1121. DOI.
Hoffmann, M., Marques, H. G., Hernandez Arieta, A., Sumioka, H., Lungarella, M., & Pfeifer, R. (2010). Body schema in robotics: A review. IEEE Transactions on Autonomous Mental Development, 2(4), 304–324. DOI.
Hu, Y., et al. (2025). Egocentric visual self-modeling for autonomous robot dynamics prediction and adaptation. npj Robotics. DOI.