Does plan recognition require a complete action sequence?

No. Many methods are designed to infer hypotheses from partial observations.

Can one person pursue several plans at once?

Yes. Interleaved plans are a recognized challenge and require an appropriate model.

Is plan recognition always probabilistic?

No. Logical, planning-based and heuristic approaches also exist, although each needs a way to manage ambiguous evidence.

AI Glossary / Graphs and symbolic AI / Intermediate

Plan recognition

Definition

Plan recognition is the task of inferring an agent's possible plans from observations of its behavior and relevant context. It seeks an explanation of how observed actions fit a larger intended course of action, often while the evidence is incomplete.

What it means

Seeing someone collect a box may fit several plans: packing an order, moving supplies or returning an item. Later actions can support one explanation and weaken others. The recognizer must reason from evidence rather than treating the first plausible interpretation as certain.

Plan recognition is related to goal recognition, but a goal describes a desired outcome while a plan describes a way of pursuing it. Several plans can serve the same goal, and similar actions can appear in plans for different goals.

It also differs from action recognition, which identifies what an observed action is. Knowing that a person opened a cabinet does not by itself establish why they did so or what sequence they intend to follow.

How it works

Specify the observations and candidate explanatory model. Some approaches use a library of known plans; others use a planning domain to reason about actions, goals and possible action sequences. Probabilistic approaches can compare hypotheses under an observation and execution model.

As evidence arrives, the recognizer updates which plans remain plausible. It may need to account for unobserved actions, alternative orderings, interruptions or several plans being pursued at once. Context can affect which plans are likely to have been adopted.

Assumptions about rationality matter. A method that expects an agent to follow an efficient plan can misinterpret mistakes, exploration or unusual preferences. Likewise, an incomplete candidate set can force every observation into an inappropriate explanation.

The output should reflect uncertainty and scope. It may be a ranked set of hypotheses, a probability distribution or a partial plan explanation, rather than a definitive statement about a person's private intention.

Three practical examples

1. Collaborative assistance: A robot infers which assembly step a person may be preparing for, then offers a relevant tool while allowing the person to correct the interpretation.

2. Workflow support: An assistant uses a sequence of application actions to suggest the next step in a known procedure. It distinguishes a suggestion from authority to carry out that step.

3. Simulation analysis: Researchers infer which strategy or task plan a software agent is following from logged actions, while testing how missing observations affect the result.

Think of it this way

Think of reading the opening moves of an unfinished recipe. Gathering flour and a bowl narrows the possibilities, but does not uniquely identify the intended dish. Additional steps and context change the explanation.

The analogy does not make intention directly observable. The recognizer uses a model of possible plans and observations, and the real activity may lie outside that model or change midway.

Worked example

A toy recognizer considers two mutually exclusive plans: pack an order, with prior probability 0.6, and restock a shelf, with prior probability 0.4. It observes an action of collecting an empty shipping carton.

Assume that this observation has probability 0.8 under the packing plan and 0.2 under the restocking plan. The unnormalized hypothesis weights are 0.6 × 0.8 = 0.48 and 0.4 × 0.2 = 0.08.

Normalizing gives a packing probability of 0.48 / 0.56, approximately 0.857, and a restocking probability of approximately 0.143. Packing becomes the stronger hypothesis, but is not certain.

These figures depend on the two-plan candidate set and the invented likelihoods. If the person is collecting the carton for a third purpose, the normalized result can be confidently misleading. Also, a later observation must be evaluated with an appropriate conditional model; blindly multiplying correlated evidence as if independent can exaggerate confidence.

Illustrative case study

In a fictional project, Willow Assembly builds an assistant that anticipates a worker's next task from tool selections. The prototype assumes each worker completes one standard plan before beginning another.

Testing reveals interruptions and interleaved tasks. The team expands its model, presents uncertain suggestions as alternatives and includes a simple correction mechanism. It evaluates whether assistance actually helps rather than counting only how often a top-ranked plan matches a restricted label set.

The case illustrates that plan recognition depends on an execution model. A system that recognizes tidy demonstrations may struggle when real work changes direction or combines several activities.

Common misconceptions and mistakes

Recognizing an action reveals the full plan. One action can belong to many plans.
The most likely plan must be the true one. Likelihood is conditional on the model and evidence.
A goal and a plan are interchangeable. A goal is an outcome; plans describe ways of reaching it.
Missing actions prove a plan was abandoned. The actions may simply be unobserved.
An inferred intention authorizes action on someone's behalf. Recognition and permission are separate issues.

Business use

Plan recognition can make assistance more relevant when a workflow has recognizable structure. Its value depends on whether useful support can be offered without overcommitting to uncertain interpretations.

Evaluate on realistic sequences with interruptions, mistakes, alternative action orders and plans absent from training or the library. Report how the system handles uncertainty, not only its accuracy among a fixed set of clean examples.

For human-facing applications, allow correction and avoid presenting inferred intent as a fact. A suggestion may be appropriate at a confidence level that would be insufficient for an irreversible action. Define the observation scope and assistance boundary explicitly, and assess the cost of both mistaken help and missed opportunities to help.

Check your understanding

Why should the worked recognizer avoid reporting that packing is certain?

Answer: The observation is compatible with both modeled plans, the likelihoods are assumptions, and the candidate set may omit the person's actual purpose.

Related terms

Abductive reasoning
Bayesian inference
Goal recognition
action recognition
intention inference
plan library
hierarchical planning
human-robot collaboration

Learn next

Write two plans that share their first two actions but lead to different goals. Identify an observation that distinguishes them and one interruption that could confuse a recognizer. Specify what the system should report before the ambiguity is resolved.

Sources

Goldman, Geib and Miller model plan recognition probabilistically through plan execution, including interleaving and context: https://arxiv.org/abs/1301.6700

Pereira and Meneguzzi use planning landmarks with incomplete observations: https://arxiv.org/abs/1604.01277

The Bayesian calculation and Willow Assembly case are original illustrations, not reported experimental results.

Previous: Plan monitoring Next: Planar (CHW) layout