Is a camera image a percept?

It can be, when the image is the input exposed to the agent at an observation step.

Is an inferred object label also a percept?

It can be if the agent's sensor interface supplies labels. Within a different architecture, the label may instead be an internal inference.

Can one percept contain multiple readings?

Yes. A percept can be a structured combination of observations rather than a single scalar.

AI Glossary / Graphs and symbolic AI / Beginner

Percept

Definition

A percept is the input an agent receives from its sensors at a particular observation step. It represents what is presented to the agent, which may be incomplete or noisy, rather than the entire true state of its environment.

What it means

For a robot, a percept might contain a camera image, a distance reading and a battery indication. For a software agent, it might be a newly received message, a retrieved page or the result of a database query. Sensors need not be physical devices.

The boundary matters. A temperature reading is an observation; the conclusion that a cooling system has failed is an interpretation. A developer can place processing inside a sensor interface, so a percept can already contain detected objects or other structured information. The chosen architecture determines the level of representation.

A single percept is also different from the percept sequence: the accumulated history of observations available to an agent.

How it works

First, define the observation interface. Specify what fields arrive, when they arrive, their units and how unavailable readings are represented. An observation step may combine several sensor channels rather than correspond to one electrical measurement.

Next, the environment and sensor process produce an input. The agent receives that percept and may combine it with memory, a model and knowledge of earlier actions. It then selects an action through its decision process.

Receiving a percept does not guarantee that the information is correct or sufficient. A disconnected sensor can supply an explicit failure indication, while an undetected fault may supply a plausible but wrong value. A partially observable task can produce the same percept in several different underlying states.

Logging the original observation separately from later interpretation helps explain how a decision was reached.

Three practical examples

1. Warehouse robotics: A navigation agent receives a range scan showing distances in several directions. Objects outside the scanner's coverage remain unobserved, even if its local map contains earlier estimates of them.

2. Document handling: An agent receives extracted text and a confidence indicator from an optical character recognition service. These are its inputs at that interface; the original page may contain information the extraction missed.

3. Home automation: A controller receives a room-temperature reading and its timestamp. The timestamp helps distinguish a current reading from an old value that has been repeated by a failed connection.

Think of it this way

Think of a percept as one delivery to an observation desk. The package may contain several reports, but it does not contain everything happening outside the building. The desk's conclusions depend on the delivery plus records already held.

The analogy stops at the delivery boundary: a machine percept can be a continuous-valued vector or image, and its production may already involve substantial computation.

Worked example

Consider a toy inspection agent whose percept is a pair: observed location and observed package presence. At one step it receives (Station A, Present).

There are four possible underlying situations consistent with a perfect local presence sensor: the package is present at A, while Station B is either occupied or empty, and the distant loading gate is either open or closed. That gives 2 × 2 = 4 possibilities outside the local observation.

The percept establishes package presence at A under the stated sensor assumption. It does not establish the condition of B or the gate. If the next action requires an open gate, the agent needs additional information or a policy that accounts for that uncertainty.

If the local sensor can misreport presence, even the conclusion about A becomes uncertain. The observation value itself has not changed; the model of its reliability has.

Illustrative case study

In a fictional project, Cedar Inspection builds a dashboard that labels every incoming camera detection as a confirmed inventory fact. During review, staff find that a partly obscured carton sometimes produces an uncertain object detection.

The team separates the received detection, its confidence metadata and the inventory belief inferred from it. The agent can request another view when the available percept does not support a consequential update.

The lesson is an architectural distinction, not a claimed accuracy improvement: recording what was observed prevents a tentative interpretation from silently becoming an unquestioned environmental fact.

Common misconceptions and mistakes

A percept is the complete environment state. It may expose only a small part of that state.
A percept must come from a physical sensor. Software interfaces can supply an agent's observations.
Every field in a percept is raw data. Sensor interfaces may return processed information.
Missing input means a negative observation. An unavailable package detector is different from a reliable detector reporting no package.
Identical percepts require identical actions. An agent using memory can act differently after different histories.

Business use

Clear percept definitions help teams agree on the evidence available to an automated decision. This is useful when integrating sensors, document services, monitoring feeds or external applications.

For each input, record provenance, timing, units and known failure modes. Test stale, incomplete and contradictory observations as well as ordinary values. If an action depends on a fact that no observation can establish, the task design needs an uncertainty-handling strategy.

An audit should distinguish received information from inferred state and selected action. That separation makes it easier to diagnose whether a problem began in sensing, interpretation or decision logic, without assuming every bad outcome was a model failure.

Check your understanding

Does the inspection agent's percept establish that the distant loading gate is open?

Answer: No. The defined percept reports only local location and package presence. Both gate conditions remain compatible with it.

Related terms

Agent
Percept sequence
Internal state
Belief state
Perception
sensor
observation
partial observability

Learn next

Choose one automated workflow and write its observation interface. For every field, distinguish received value, reliability assumption and inferred conclusion. Add one example where two different environmental states produce the same percept.

Sources

Russell and Norvig's agent chapter illustrates percept inputs, sensor interfaces, history-dependent agent functions and partial observability: https://aima.cs.berkeley.edu/slides-pdf/chapter02.pdf

The inspection calculation and Cedar Inspection case are original explanatory examples. Their sensor assumptions are explicit and do not describe measured hardware performance.

Previous: Percentile Bootstrap Confidence Interval Next: Percept schema