The Sense-Estimate-Plan-Act-Check Loop

Begin With a Failure

A delivery robot sees an obstacle, turns left, and still clips a box. Saying “the AI failed” is not useful enough to repair it. Engineers need to locate the failure: Was the box sensed? Was its position estimated incorrectly? Did the planner choose an unsafe path? Did the wheels execute the command? Did the robot check the result?

That investigation follows a reusable loop:

goal -> sense -> estimate -> plan -> act -> check -> update

1. Goal: Define Success and Boundaries

The goal says what outcome matters. “Reach loading bay B” is a goal. Constraints refine it: stay inside the marked lane, keep 0.5 metres from people, and stop when confidence is too low. A robot cannot judge success unless designers make the objective and limits observable.

2. Sense: Collect Measurements

Sensors produce measurements, not truth. A camera produces pixels. A lidar returns distances. Wheel encoders report rotation. Measurements can be noisy, delayed, blocked, or ambiguous.

For the delivery robot, a depth camera might report a cluster of points two metres ahead. That is evidence, but it is not yet a complete claim about the world.

3. Estimate: Build the Best Current State

Estimation turns measurements into a working belief: “A stationary box is two metres ahead and slightly right; the robot is moving at 0.4 metres per second.” The estimate may combine several sensors and always contains uncertainty.

This distinction is crucial:

• measurement: what a sensor reported;
• state estimate: what the robot currently believes about itself and the environment.

4. Plan: Choose an Action Sequence

Planning compares possible actions against the goal and constraints. The robot may slow down, steer around the box, stop, or request help. A plan is only as good as the state and assumptions it receives.

5. Act: Affect the Physical World

Controllers translate a planned motion into actuator commands. Motors, brakes, joints, and grippers have limits. Wheels can slip; batteries sag; payloads change stopping distance. A correct plan can still fail during execution.

6. Check: Compare Expected and Observed Results

After acting, the robot senses again. It compares the expected result with evidence. Did it turn far enough? Is the obstacle farther away? Is the goal closer? A mismatch triggers correction, replanning, stopping, or human escalation.

Without checking, the chain becomes open loop: command once and hope.

Worked Trace: Avoiding the Box

The table shows why failures propagate. A late or incorrect estimate can produce a reasonable plan for the wrong world.

Watch: The Robot Evidence Loop

Watch How Robots Think: Sense, Estimate, Plan, Act & Learn. Before pressing play, write the six loop stages from memory.

While watching, capture one concrete item for each stage:

After watching, explain where the human goal and safety limits enter the loop. Then answer: if the camera measured the box correctly but its stored position was stale, which is the earliest broken link?

Use the accompanying vertical infographic as a reference, but reconstruct the loop without looking before attempting the retrieval check.

The Loop Runs at Several Speeds

A motor controller may check velocity hundreds of times per second. Navigation may replan every second. A human supervisor may review a mission after several minutes. These nested loops should exchange clear state, confidence, and stop conditions.

Retrieval Check

1. Why is a sensor measurement not the same as a state estimate?
2. Give one example of a correct plan failing during action.
3. What evidence should the check stage compare?
4. Where would you place a human escalation rule in the loop?

Key Takeaway

Intelligent robot behavior is not one mysterious decision. It is a chain of evidence and responsibility. Trace the goal, measurement, estimate, plan, physical action, and feedback separately; then repair the earliest broken link.