Detection Is Not Understanding

A perception model can output a label and confidence score without understanding the object the way a person does.

That matters when a robot uses the output to act.

A detection record

A vision system might return:

{
  "label": "box",
  "confidence": 0.91,
  "center_x": 412,
  "center_y": 287,
  "width": 138,
  "height": 121
}

The record is useful. It is not proof.

The system may still be wrong about:

• the label;
• the real-world position;
• whether the object is reachable;
• whether another object blocks the path;
• whether 0.91 is well calibrated.

False positives and false negatives

A false positive reports something that is not actually present.

A false negative misses something that is present.

Robotics cares about the consequences.

Missing a decorative marker may be harmless. Missing a person in a motion zone is not.

The acceptable error rate depends on the task and safety architecture.

Confidence is not probability of safety

A high model confidence does not mean:

91% chance this action is safe.

It describes the model's output under its own scoring system. The control system still needs independent rules.

Decision boundary

Suppose a robot should pick up a detected object.

Before acting, it might require:

• detection above a threshold;
• estimated pose inside reachable workspace;
• object observed across several frames;
• no stop condition;
• motion zone clear.

The perception result contributes to the decision. It should not automatically own the actuator.

Build a perception-to-action table

Choose a camera-based robot task.

For each step, record:

vision output → interpretation → decision rule → actuator consequence

Then add one false positive and one false negative case.

Explain which failure is more serious and why.

Detection is one stage in a larger decision system

Suppose a model outputs:

{
  "label": "person",
  "confidence": 0.82,
  "box": [310, 120, 480, 420]
}

The model produced a classification-like result and a location. It did not decide whether the robot should stop, slow down, speak, reroute, or ignore the detection.

Those are policy decisions made elsewhere.

image
  ↓
perception model
  ↓
detection result
  ↓
decision rule / policy
  ↓
robot action

Separating those stages is critical. A threshold such as confidence > 0.70 is a human-selected rule, not something the camera discovered about safety.

False positives and false negatives have different costs

For a safety-related stop detector:

• a false positive may stop the robot unnecessarily;
• a false negative may allow motion when a hazard is present.

The correct threshold depends on consequences, redundancy, operating speed, and other safety controls.

Do not describe a perception system with one accuracy number alone. Ask what kinds of mistakes it makes and what the rest of the robot does when those mistakes occur.