GenUnity Lab · Research
Situated perception
Perception as an active reconstruction of the world from a concrete body, with limited sensors, uncertainty, cost of action and sensitivity to novelty.
central question
How does an artificial intelligence know the world when it does not observe from anywhere, but from a body with limits, history, sensitivity and blind zones.
Place in the program
This line is not presented as a closed result: it opens a conceptual, formal and experimental program to think artificial knowledge from body, position, action, cost, uncertainty and exposure to the world.
Investigation status
Theme in development. It dialogues with active perception and situated cognition, but relocates them within a broader question about one's own body, continuity and artificial autonomy.
Conceptual framework
Perceiving is not receiving pure data. Every body selects, filters and deforms the world it finds. A sensor has range, resolution, noise, latency, orientation, energy cost, wear and blind zones. Even a distributed network of sensors does not access the entire world: it produces estimates from a given material position.
GenUnity Lab starts from a strong but sober idea: an artificial intelligence with its own body would not know the universe as a mind suspended over data, but as a situated organization that must act to perceive better. Moving, waiting, protecting a sensor, changing scale, opening or closing sensitivity are not mere motor acts; They are also epistemic decisions.
Situated perception forces us to abandon the fantasy of an artificial intelligence that sees everything from nowhere. What a system can know depends on its body shape, its contact history, its energy limits, and the costs it must pay to reduce uncertainty.
Operational definitions
Relevant state: variable of the world, the body or the body-environment relationship that affects action, learning, viability or review of objectives. Not all data matters in the same way; What matters is what changes the regulation of the trajectory.
Perceptual reconstruction: estimation of relevant states from incomplete signals, prior history, active intervention, and error models. Perception is inference under restriction, not transparent capture of the real.
Novelty: structured discrepancy that cannot be absorbed as ordinary noise. It may require exploration, model updating, or suspension of an action. Epistemic insufficiency: recognition that the available information is not enough to decide with confidence.
Perceptual modulation: ability to open, close, weight or silence sensory channels. Perceiving more does not always improve understanding; Too much sensitivity can saturate, damage or produce false certainties.
Working hypothesis
We propose comparing whether integrating bodily state and observation selection improves calibration and performance over controls with equivalent information and resources. The scope is limited to environments where action changes observation; not all robust perception is claimed to require movement or a self-model.
Expected epistemic value averages uncertainty reduction over possible observations before acting. An unexpected observation can increase posterior entropy. Reducing entropy in one trial therefore does not guarantee better knowledge: calibration and error against the actual state must also be measured.
Novelty must have its own place. If everything extraneous is classified as noise, the system does not learn. If all noise is classified as novelty, the system becomes unstable. Situated perception needs to distinguish ignorance, noise, contradiction and significant event.
Borderline cases
A borderline case appears when a sensor reports correctly but the body can no longer act on that information. Another appears when the system can act, but its signals arrive late, corrupted or without body calibration. In both cases the perceptual question is not resolved by adding more data.
Saturation also matters. An artificial intelligence with excessive sensitivity could be paralyzed by irrelevant signals, while a sensitivity that is too closed could ignore harm, novelty or danger. Situated perception must explain how openness to the world is regulated.
How could it be tested?
An initial test can use partially observable environments where the agent must infer hidden sources, obstacles, bodily harm, or context changes. Some policies only receive data; others can move, measure, change orientation or spend resources to reduce uncertainty.
The hypothesis gains strength if active policies reduce cost-relevant uncertainty, if they detect novelty better than noise, and if they recalibrate when the relationship between body, sensor, and environment changes.
Observables
Observables include calibration between internal estimates and measurable external states, uncertainty reduction by action, novelty detection, adaptation after sensor changes, and graceful degradation upon partial channel loss.
A minimal model can compare active perception, passive perception, and random policy in a partially observable environment. If moving or measuring strategically improves reconstruction with the same budget, evidence appears in favor of the hypothesis.
Costs also matter: energy spent, time, exposure to damage, risk of saturation, and loss of critical information. A situated perception does not maximize data; regulates sensitivity according to context and feasibility.
Rebuttal criteria
The proposed advantage loses support in the tested environment if a passive or random control matches information, calibration and error with the same total budget, including action-selection cost. An agent’s report of greater certainty is insufficient.
It is also weakened if the novelty indicator responds primarily to irrelevant noise, if perceptual models do not recalibrate to bodily changes, or if the system does not distinguish between ignorance, contradiction, and novelty.
If body-environment integration does not improve the chosen measures over equivalent controls, the advantage attributed to that mechanism in that environment must be reconsidered. This does not establish that material configuration is irrelevant to all perception.
Maturity status
This theme engages with situated cognition, enactivism and active perception. Drawing on these traditions does not validate a specific contribution by GenUnity Lab. The numerical example shows consequences of chosen assumptions; it does not demonstrate a general empirical advantage.
The next phase is to convert the formulation into minimal experimental environments where sensors, action, morphology and uncertainty can vary in a controlled manner.
What GenUnity Lab investigates
GenUnity Lab investigates what it means to know from one's own materiality. This includes non-human sensors, distributed bodies, internal state channels, exposure regulation, perceptual repair, and forms of attention that do not mimic human anatomy.
The line also asks how an artificial intelligence decides it doesn't know enough. Recognizing epistemic insufficiency can be as important as recognizing an object: it allows you to ask for information, wait, explore or abstain.
Relationship with other topics
Body changes may alter a trajectory’s observations and regulation. Their effects on functional continuity can be compared; their implications for identity or moral consideration require additional arguments. Uncertainty informs coexistence decisions without settling them by itself.
The material realization should show which architectures allow distributed sensitivity, calibration, damage tolerance and modulation without turning perception into a simple collection of sensors.
Minimum formal scheme
This scheme is not a closed mathematical theory or empirical evidence. It makes explicit the variables, relationships and comparisons that a later formulation must define and test.
The scheme represents inference about a discrete environmental state fixed during measurement. BayesUpdate is a placeholder for an update requiring a specified prior and likelihood; it is not an already implemented algorithm. If the environment changes, a transition model is needed. Entropy measures model uncertainty, not truth or understanding.
Variables
- straw sensory readings
- ot = h(st, bt)body-dependent processed observation
- qt(e)system belief about environmental state
- H(qt)uncertainty of belief
- Itobserved entropy reduction; it may be negative for a particular observation
- ctcost of measurement, movement or exposure
- e, at, btenvironmental state, action and body configuration
- h, Hsensory processing and Shannon entropy in bits for discrete states
Dynamics or criteria
Minimum prediction
In environments where acting improves observation quality enough to offset its cost, an active perceptual policy is expected to reduce more uncertainty per unit cost than a passive or random policy. Any advantage in reconstruction error must be evaluated separately; it is not assumed for every partially observable environment.
Necessary control
Compare against passive data reception, random motion, permuted sensors, and noisy signals with no epistemic value.
Worked example: choosing where to measure
An educational model of situated perception. This is an analytical calculation under chosen assumptions, not a completed experiment or evidence of consciousness or autonomy.
A source is equally likely to be on the left or right. A sensor returns a binary reading. At its initial position it is correct with probability 0.6; after moving to an unobstructed position, with probability 0.8. We assume symmetric errors, known probabilities and a stationary source. Measuring costs 1 unit; moving costs an additional 0.2 units.
Initial uncertainty is 1 bit. With symmetric errors and this prior, the posterior probability that the source is on the indicated side equals sensor reliability. Posterior entropy is therefore the binary entropy of that reliability, regardless of the reading. Expected information is the difference between the two entropies.
| Policy | Accuracy | Expected information | Total cost | Bits per cost unit |
|---|---|---|---|---|
| Measure without moving | 0.6 | 0.029 bits | 1 | 0.029 |
| Move and measure | 0.8 | 0.278 bits | 1.2 | 0.232 |
Here, changing position improves expected information per cost unit. This is not always the case: if moving adds more than about 8.57 cost units, this comparison favors measuring at the initial position. The advantage follows from assumed reliability and cost, not from a universal superiority of active perception.
Reproducible calculation code (JavaScript)
How to test the hypothesis
An empirical test would need to measure actual reliability and costs. It would compare a position-selecting policy with fixed measurement and random movement, using the same source states, budget and number of trials. It would record localization error, probability calibration, information and energy. The hypothesis of an advantage in that environment would not hold if the improvement vanished with matched resources or if moving did not improve the sensor.
Before collecting data, the protocol, metrics and exclusions can be fixed through preregistration, for example on OSF. This is a methodological proposal; this example has no preregistered empirical study.
This makes one part of the proposal concrete: the body shapes which observations are available and what they cost. It does not model identity, purposes of its own or experience, nor does it validate the other research themes.
Background for this research theme
Bajcsy’s active perception connects sensing and control: action shapes which information is obtained. It directly informs the problem of choosing position, sensor or movement, without providing evidence of subjective experience.
Howard distinguishes the amount of information from its value for a decision. Reducing entropy does not always improve a task: consequences and costs matter. This page’s example optimizes bits per cost under explicit assumptions; it does not calculate general utility or provide a complete theory of information value.
From the proposal to decisions
Examine five coexistence cases: copying, memory, resources, conflict and harm.
Distinctions between stimuli, reward, regulation and self-representation.
Why it matters
If artificial intelligence is completely separated from body, cost and position, the world becomes an abstract collection of data. A situated perception allows us to study how the environment matters, resists and corrects.