# Learning from Labeled Examples — complete non-video alternative

All objects, labels and values are synthetic. This original animation does not demonstrate a deployed robot or safety-rated controller.

## 0:00 — Perfect training. Wrong bin.

**Visual mechanism:** A wheeled sorting cart carries a narrow parcel toward a fork with N and W bins. The proposal points toward W; the cart stops before a visible hold gate.

**Complete narration:** This robot sorter gets every training package right. Tomorrow, a narrow package heads toward the wide bin. Did the robot forget? Not necessarily. A changed measurement can break a learned rule without changing its code.

## 0:15 — Learn from labeled pairs

**Visual mechanism:** A camera scans a parcel; a width bracket becomes x=4. A separate inspection card supplies label N. Both arrows feed a training pair, not a motor.

**Complete narration:** Each example pairs an input feature with a reference label. Here, the feature is a simulated camera width, and the label is narrow or wide. People define the task and check the labels. The model does not discover what our categories ought to mean.

## 0:33 — Move a rule. Count errors.

**Visual mechanism:** Six labeled points on a vertical width axis. The candidate threshold moves from 3.5 to 5; C changes prediction and the training-error counter moves from one to zero.

**Complete narration:** Our model predicts narrow below a threshold and wide otherwise. Try the candidate thresholds on six training examples. At three point five, one label is wrong. At five, none are wrong. Fitting chooses five from the data. We engineered the model family; the threshold was learned.

## 0:55 — Three sets. Different jobs.

**Visual mechanism:** Three stacked trays label training, validation and test roles. A second frame A2 approaches the test tray, meets a rejection mark and returns to the training group. The instruction is to freeze choices before final testing.

**Complete narration:** Training fits the rule. Validation guides development choices. Freeze those choices before the final test. Keep related frames of one object together when testing new objects. Otherwise, a familiar package can leak into the test. Even two correct unseen examples cannot prove reliability.

## 1:17 — Inference does not retrain

**Visual mechanism:** Two unseen objects pass a camera one at a time, a fixed t=5 rule produces N and W, and comparison cards record 2/2 nominal test.

**Complete narration:** With the threshold frozen at five, a new width of three point five predicts narrow. Six point five predicts wide. Both match this tiny nominal test. Report two out of two and the conditions, not a guarantee. The labels check the predictions; they are not inputs to inference.

## 1:38 — Same object. Different x.

**Visual mechanism:** The same narrow object is measured; a +3 bias pulse changes its readout from 3.5 to 6.5. The proposal changes N to W, and a reference N card exposes the error.

**Complete narration:** Now add three to each measurement as a separate stress test. The narrow object reads six point five and is predicted wide. The labels and threshold stayed fixed, but accuracy fell to one out of two. More copies of yesterday's data do not represent this changed condition.

## 1:58 — Does a prediction permit motion?

**Visual mechanism:** Prediction W card feeds a closed dispatch gate. Sensor checked = no. Cart remains stationary while four seconds of untimed-friendly retrieval countdown is shown.

**Complete narration:** The classifier proposes wide, but the sensor condition has not been checked. Should the robot dispatch a new movement? Name the missing evidence before you continue.

**Untimed retrieval prompt:** Pause here and give your own answer before reading the next section. The video allows four seconds, but this alternative has no time limit.

## 2:14 — Prediction is not permission

**Visual mechanism:** Sensor and route checks remain separate from the model output. A held cart and crossed dispatch arrow clarify that this is only a simulation policy. Closing three-step summary is labeled fit, test, evidence.

**Complete narration:** No. Keep new dispatch on hold while the sensor condition is unresolved. A class label does not verify a route or make motion safe. Try the sorter lab and paper alternative in the AI Robotics course. Fit the rule, protect the test, and explain a failure with evidence.

The camera feature is in model units. N/W mean narrow/wide. The fitted rule is N if x < 5, W otherwise. The reference labels are independent of the injected measurement bias. A category prediction is not motion authority.