Is a percept schema the same as a percept?

No. The schema models possible observations and their conditions; a percept is an observation received during operation.

Does observing a property create it?

Not generally. In the basic planning model, sensing reveals information about state rather than changing that property.

Does the schema handle noisy sensors automatically?

No. Uncertainty and failure behavior need an appropriate model beyond a simple deterministic observation rule.

AI Glossary / Data engineering and architecture / Intermediate

Percept schema

Definition

A percept schema, in AI planning, is a structured description of what an agent can observe and under which conditions that observation becomes available. It gives a planner a model of sensing, especially when the environment is only partially observable. It is not a database schema, despite the word schema, and it should be distinguished from the broader perceptual-schema terminology used in some robot-control architectures.

What it means

An agent may know that a package has a label but be unable to read it until the label faces the camera. A percept schema represents that relationship between conditions and obtainable information. The planner can then include actions that make observation possible instead of assuming it already knows every relevant fact. The schema describes the availability of knowledge; the actual percept is supplied when sensing occurs during execution.

How it works

Specify the property or fluent that can be observed and the preconditions required for sensing it. The planner reasons about actions that establish those preconditions, such as moving into view or opening a cover. It then uses the observation to update its belief about the world and, where needed, choose a contingent branch. A simple symbolic model may assume accurate observations when conditions hold. Real sensor noise, ambiguous results, and failures require additional modeling rather than being solved merely by writing a percept schema.

Three practical examples

Inspection planning: an agent knows that a serial label can be read only when the relevant surface is visible.
Robotic handling: a planner includes an authorized inspection step before choosing between actions that depend on an unknown property.
Software-agent design: a comparable observation contract states which tool response can reveal a status and what access or context is needed, while distinguishing that analogy from a formal implementation of the planning language.

Think of it this way

Think of a rule saying that you can read a notice only after reaching the noticeboard and facing it. The rule tells you how to obtain information; it does not tell you in advance what the notice says. The analogy is limited because planning formalisms specify precise state conditions and belief updates. Real perception may fail or be uncertain even when you appear to be in the right position.

Worked example

A toy warehouse agent must place a sealed package in the correct area based on whether its label says fragile. Initially it knows the label is on the rear face but does not know the label value. A percept schema says that Fragile(package) becomes observable when LabelVisible(package) holds and the scanner is ready. The plan first rotates the package using an allowed handling action, then observes. If fragile is observed, it follows the protected-placement branch; otherwise it follows the ordinary-placement branch. The observation does not make the package fragile or non-fragile; it changes what the agent knows about an existing property. If the scanner returns unreadable, that result needs an explicit fallback rather than being treated as false. This is a conceptual example, not a validated robot-control procedure.

Illustrative case study

Fictional example: Hazel Automation prototypes an inventory assistant that assumes every barcode can be read immediately. Tests reveal that occluded labels are being treated as absent items. The team revises its observation model to distinguish a visible readable label, an unreadable observation, and no observation yet. It then designs review or repositioning behavior for the unresolved cases. This invented case illustrates better sensing assumptions; it does not claim a measured recognition improvement or physical safety certification.

Common misconceptions and mistakes

Do not confuse a percept schema with a particular sensor reading or with an action that changes the world. Sensing can update knowledge without changing the observed property. Do not assume that failing to observe a fact proves it false. Avoid calling every data format a percept schema simply because it holds sensor output. Also distinguish planning's symbolic sensing model from a perceptual module that processes raw images or signals; the two can work together but describe different responsibilities.

Business use

The concept is useful when a workflow depends on information that becomes available only after specific steps. List unknown facts, identify how they can be observed, and state observation preconditions honestly. Include failure and uncertainty outcomes before allowing downstream decisions to rely on them. For AI agents, this encourages a clear distinction between retrieved evidence, inferred state, and information not yet obtained. It also helps reviewers see whether a plan assumes access or sensing capability that the deployed system does not actually possess.

Check your understanding

An agent cannot see a label because it faces away. Can it conclude the label's statement is false?

Answer: No. The observation precondition is unmet. The property remains unknown until suitable evidence is obtained or another justified inference establishes it.

Related terms

Percept
Belief state
Action schema
Precondition
partial observability
sensing action
contingent planning
observation model
sensor model
perceptual schema

Learn next

describe a small task with one unknown property and write the conditions under which it can be observed. Build branches for each observation result, including failure to obtain a result. Separate actions that alter the world from steps that only improve the agent's knowledge.

Sources

Sources: Michael Hahsler, Automated Planning course slides, https://mhahsler.github.io/CS7320-AI/slides/11_Automated_Planning.pdf ; Carnegie Mellon University, Planning and Acting under Uncertainty handout, https://www.cs.cmu.edu/~mmv/planning/handouts/PlanningUncertainty.pptx.pdf ; AIMA authors' index, https://aima.cs.berkeley.edu/aima-index.html . Primary instructor excerpts consulted for percept schemas and partial observability; the authors' index confirms the term. Direct access to the Hahsler PDF failed after redirection, so full-slide review remains pending. Examples are original and conceptual.

Previous: Percept Next: Percept sequence