T6 — Belief Under Uncertainty

Learning objectives

After this lesson, you should be able to:

• Explain why robots operate under uncertainty.
• Distinguish true state, observation, estimate and belief.
• Explain what a belief represents.
• Understand confidence and competing hypotheses.
• Explain how new observations update belief.
• Understand why uncertainty affects planning and action.
• Explain the connection between belief and partial observability.
• Recognize when a robot should gather more information instead of acting immediately.

The key mental model is a loop, and the whole lesson fills it in:

The one-sentence idea

Because a robot cannot directly know the true state of the world, it must maintain and update a belief about what is likely to be true.

A robot rarely knows the exact state of the world. It keeps a belief about what may be true, updates it as observations arrive, and decides from that.

Why robots are uncertain

Uncertainty is fundamental to physical systems, not a sign of a poorly built robot. It comes from three directions:

The sensing

Sensor noise, occlusion, a limited field of view, missing observations, and ambiguous observations (is that a bottle or a vase?).

The world

Moving objects, and observations that arrive delayed.

The robot itself

Localization uncertainty, imperfect models of how it moves, and actions that do not produce exactly the expected result.

Take a camera that reports something that might be a red bottle, with part of it hidden. The robot should not treat “red bottle detected” as the same thing as “the robot knows with certainty that the red bottle is at this location.” The first is evidence. The second would be knowledge, and the robot does not have it.

True state, observation, estimate, belief

These four ideas are foundational to Physical AI, so it is worth seeing them side by side. Use the red bottle:

Two more words travel with these. Uncertainty is how unsure the robot is about the state. Confidence is a number or label saying how far the robot, or one of its components, trusts a particular estimate or detection.

What is a belief?

A belief answers one question: what does the robot currently think could be true? Suppose a robot is unsure where it is:

These numbers are not necessarily perfect probabilities in every engineering system. They are an intuitive way to represent uncertainty. One more point deserves care: a belief does not mean the world itself is uncertain. The physical world has a particular state, and the robot is in exactly one place. It is the robot that is uncertain about that state.

Single vs multiple hypotheses

Compare “I believe the object is at (2.1, 1.5)” with “the object could be at several locations.” The first commits to a single estimate. The second keeps multiple hypotheses alive.

Keeping several is useful when the evidence does not yet decide between them. In localization, a robot in a building with two identical corridors might see a wall that fits both, so it keeps both positions in its belief until something tells them apart. For object detection, two detections a meter apart might be one object seen twice, or two objects. Committing too early to one guess is how a robot confidently goes the wrong way.

Confidence is not truth

High confidence does not guarantee correctness. A vision model can confidently detect the wrong object. Localization can be confidently wrong. A stale observation can look reliable even though the world has changed since. Confidence measures how strongly the system holds a belief, not whether the belief is right.

Confidence should influence decisions, not replace verification.

This is the idea behind Lab 09: a confident success report is still just a report until something checks it.

Belief update

The basic move is simple: take the old belief, add a new observation, and get an updated belief.

1. The robot starts with some belief.
2. It receives a new observation.
3. It compares the observation with each possible state.
4. It raises belief in states consistent with the observation.
5. It lowers belief in states that are inconsistent with it.
6. It keeps acting with the updated belief.

Here the robot sees a kitchen doorway ahead:

Kitchen did not jump to 100%. The doorway is evidence, not proof, so the other possibilities shrink but remain.

Prediction and correction

Observations do not arrive continuously, so between them the robot predicts how the world has changed, and then a new observation corrects that prediction.

For localization: the previous belief says the robot is near the kitchen. It drives forward, so the prediction is that it is now closer to the kitchen entrance. Then a sensor detects a door or a landmark, and the robot corrects its belief about where it is. Kalman filters, particle filters and Bayesian filtering are all ways of doing this general thing well. Their algorithms belong in the Localization & SLAM track (coming soon); for a hands-on first look at drift and correction, see Where Am I?

Partial observability

If a robot cannot observe everything about the world, it cannot determine the complete state directly. So it keeps a belief over the possible states. A red bottle is hidden behind an object. The robot does not know whether:

• the bottle is still on the table,
• someone moved it,
• it fell, or
• it is just outside the camera’s field of view.

Each is a different state of the world, and each calls for a different next step. The robot must reason about all of them. This is exactly the situation Lab 03, Partial Observability (in development), will let you work with.

Belief and the world model

In T5 the world model held facts such as:

at(red_bottle, kitchen_table)

A world model does not have to hold only true-or-false facts. It might instead record belief(at(red_bottle, kitchen_table)) = high, or keep several candidate locations. It can represent:

• Known facts
• Uncertain facts
• Confidence
• Competing hypotheses
• Stale information

Belief is the part of the world model that says how much each fact should be trusted.

Belief changes over time

At 10:01 the robot should not carry on using the old world model. The bottle may have moved, been picked up, be hidden, or the detector may simply have failed. The belief has to change, and that naturally leads to reasoning and replanning.

Acting under uncertainty

A robot does not always need certainty before acting. There is an engineering tradeoff between two choices:

Good Physical AI does not eliminate uncertainty; it manages uncertainty.

Information-gathering actions

Some actions are valuable mainly because they reduce uncertainty:

• Move the camera
• Look around an obstacle
• Scan the environment
• Check the gripper
• Inspect an object
• Ask the user
• Re-observe the scene

None of these moves the bottle. They change what the robot knows, which changes what it should do next. A later lesson, T16 (Act, Inspect or Ask), turns this into a decision rule; here we only need the idea.

Example: finding the red bottle

The user says, “Bring me the red bottle.” The robot observes the kitchen and forms an initial belief about where the bottle is:

The robot approaches the table, but the camera cannot clearly see the bottle. It should neither assume failure nor assume the bottle is simply not there. It updates its belief, and then it has options:

1. Move the camera.
2. Inspect another angle.
3. Check the cabinet.
4. Search another location.
5. Ask the user, if the ambiguity remains.

Each step is the loop from the top of the lesson: observation, belief update, decision, action, and a new observation that updates the belief again.

Why belief matters for planning

A planner should work from the robot’s current belief, not pretend it has perfect knowledge. That can change the plan itself:

• If the robot believes a door is probably closed, it may plan to open it.
• If it is unsure whether the door exists, it may inspect first.
• If it is unsure which object the user means, it may ask.

Same goal, three different plans, because the beliefs differ.

Common misconceptions

Engineering takeaways

1. Robots do not directly access the true state of the world.
2. Sensors provide observations, not certainty.
3. Belief represents what the robot currently considers possible or likely.
4. New observations should update belief.
5. Uncertainty affects planning and action.
6. Some actions exist mainly to reduce uncertainty.
7. A robot should distinguish between “I know” and “I believe.”
8. Good Physical AI continuously updates its understanding of the world.

A robot does not know the world directly. It maintains a belief about the world, updates that belief from observations, and uses it to decide what to do next.

Knowledge check

Three conceptual questions. Write an answer, then reveal the explanation. Your answers stay in your browser.

What’s next

The robot can now represent uncertain information about the world. The next challenge is turning a human command into something the robot can actually execute.

Related content

Nothing below is required to finish this lesson.

• Lesson
  T2: State, Observation and Belief

  Where belief was first introduced.

• Lesson
  T4: From Sensors to Observations

  How the observations that update belief are produced.

• Lesson
  T5: World Models and Symbolic State

  Where belief lives, inside the world model.

• Lab 03
  Partial Observability In development

  Turn noisy, partial observations into a belief.

• Track
  Robot Foundations and its Where Am I? lab

  Odometry drift and landmark correction: prediction and correction in practice.

• Track
  Perception and Localization & SLAM Coming soon

  How observations are produced, and how robots localize in depth.