Learning from Labeled Examples

The sorter labels every training package correctly. A sensor changes, and the next narrow package goes toward the wide bin. Did the robot forget—or did we ask the wrong question when testing it?

Learning outcomes

• Separate an input feature, a target label, a learned parameter and a control command.
• Fit and inspect a tiny classifier using labeled training examples.
• Plan training, validation and final testing without mixing related objects across sets.
• Explain a failure under a changed measurement and preserve explicit intervention boundaries.

Prerequisites: module 6 perception errors and 9.3 evidence versus authority. Budget: 10 minutes concepts, 10 worked example, 25 lab, 10 quiz/reflection. No robot, paid service, package installation or video is needed.

Before / During / After

Before: Predict what “six out of six training examples correct” establishes. Is it enough to send motor commands?

During: Follow a package from camera measurement through a threshold to a proposed bin. Watch the threshold move only during fitting. Notice which observations are withheld.

After: Explain why adding three to a measurement can change a prediction without changing the trained model. Complete the same trace with the paper data below.

Watch the complete Humanoid Hub tutorial. The descriptive transcript and paper alternative cover the same learning without video.

1. What is learned?

A feature is an input available when making a prediction. A label is the target supplied with a training example. In our original simulation, the input is an extracted width, x, and the target is an inspected category: N for narrow or W for wide. The model is a threshold, t: predict N when x < t; otherwise W. Inference means applying the fixed rule to another input. Google's supervised-learning introduction describes these roles.

We deliberately engineer the model family, input representation and label definitions. Fitting selects t from data. A model need not be a neural network to be learned. A rule with a manually chosen t and a rule with a fitted t can have the same inference code; the source of the parameter differs.

The scalar is not a physical ground-truth dimension: it is a fictional camera-derived measurement in model units. We assume the inspection labels are the reference for this exercise. Ask who labeled real data, using what definition, and how disagreements and ambiguous cases were handled. A model can learn an annotation mistake.

2. Fit a rule you can inspect

Try each candidate threshold and count wrong predictions. We test midpoints between successive observed values plus one outer value at each end. Ties choose the lowest threshold.

The selected t is 5. For C, 4 < 5, so predict N. For D, 6 < 5 is false, so predict W. Exactly 5 is W under our stated boundary convention. No validation or test label helped select this threshold. A constant-W baseline is only 3/6 on this balanced training set.

Retrieval pause: At threshold 3.5, which training object is wrong? Try before checking: C is predicted W despite label N.

3. Give each dataset a job

Use training for fitting. Use validation for development choices, such as whether a feature or model family is useful. Freeze those choices before final testing. Do not repeatedly change a model to improve the same test result while calling that result independent. Google's split guide explains the distinction.

These are deliberately tiny invented sets so you can check every calculation. Two successes do not measure real deployment reliability. Report the denominator and conditions, not just “100%.” If validation leads you to revise the design, that is development—not an untouched test.

Adjacent frames of the same package share information. A second frame of A is not a genuinely new object. For an unseen-object question, keep an object's frames in one split. Testing a new site or future session requires a split designed for that question. Group-aware and time-aware evaluation are discussed in scikit-learn's cross-validation guide.

If you learn a scaler or select features, learn those transformations only from training data and reuse them consistently. Computing a normalization value using the final test set leaks information into the pipeline. scikit-learn's pitfalls guide explains why fitting and transformation are different operations.

4. Change the sensor, not the labels

Freeze t = 5. As a separate stress scenario, add +3 to each nominal test measurement. Package I is still labeled N, but its reported width is now 6.5, so the model predicts W. J reports 9.5 and predicts W correctly. The result is 1/2, including an N-to-W error. The object labels and threshold did not change; the input relationship did.

This is a controlled counterexample, not a claim that every camera failure is a constant offset. More copies of the old training rows do not represent a new sensor condition. Collect and inspect relevant evidence, check preprocessing/calibration, develop with an appropriate validation set, and reserve new testing for the revised pipeline. Do not silently retrain on the failed test and reuse its score as independent evidence.

5. A prediction is not permission

Open the learning and dispatch diagram. Equivalent text: training pairs fit t=5; a new width gets a category prediction; an independent reference label evaluates it; separate sensor and route checks decide whether this simulation may propose new dispatch.

Our diagram separates data/model evaluation from a proposed bin and an explicit dispatch check. A learned category alone does not check clearance, contact, braking, people nearby or sensor validity. MIT's manipulation notes describe combining learned perception with planning and control. Our simulator holds new dispatch when the sensor condition or route is unchecked; it never controls hardware.

The hold gate demonstrates separation of responsibilities, not certified protection. As in 9.3, a software state must not be mistaken for proof of a physically safe machine. No probability of correctness is computed by this one-threshold classifier.

Misconceptions and transfer

• “Training perfection means understanding.” It only describes those labeled rows under that evaluation.
• “Two camera frames are two independent objects.” Their shared object or session may leak information.
• “More rows always help.” Repetition cannot supply absent conditions; wrong labels can reinforce errors.
• “A high class score makes an action safe.” This model has no confidence score, and prediction and motion authority are distinct.

Try the lab, then the seven-question check. Submit a threshold calculation, a split rationale, a sensor-shift counterexample and a limitation. Takeaway: define the target → fit on training data → develop honestly → test on suitable unseen cases → keep control checks explicit. Next lesson distinguishes learning from demonstrations and rewards.

Full source/rights details are included below. Full non-video mechanism: descriptive transcript.

10.1 source and rights ledger

Fresh primary-source retrieval: 2026-10-12 UTC. Original synthesis, synthetic data and diagrams; no third-party figures, footage, datasets, exam questions or source code are copied.

The supervised-learning index was inspected for orientation, not used as evidence for a particular algorithm. Two initially attempted Google Crash Course URLs returned tool errors; the working Intro-to-ML supervised page was inspected instead. No claim relies on failed retrieval.

Original example and checks

N and W are narrow/wide package categories assigned by a fictional inspection process. A simulated image extractor reports a scalar width feature in model units. Train rows A–F, validation G–H and test I–J are different object IDs. The threshold family and lowest-threshold tie policy are explicitly engineered; the fitted threshold is selected by training errors. This is supervised learning even though its model is one number.

Candidate midpoints: 1.5, 2.5, 3.5, 5, 6.5, 7.5, 8.5. Training error counts: 3, 2, 1, 0, 1, 2, 3. The constant-W baseline gets 3/6 on training and 1/2 on the balanced two-item test. Validation is 2/2 and untouched nominal test is 2/2 at threshold 5; these tiny constructed sets are arithmetic demonstrations, not empirical robot performance. A +3 sensor bias on test produces 6.5 and 9.5, both predicted W; one of two is wrong. This separate stress scenario does not replace the frozen nominal test.

The missing-label and wrong-label risks, explicit sensor/route checks, and synthetic sorter animation are original teaching constructions. No classifier confidence is computed. A HOLD decision is a simulated dispatch gate, not an emergency stop. Real perception, sensor calibration, actuation, braking, collision avoidance and safety are not validated.

Assessment alignment

Q1 feature/target; Q2 threshold fit; Q3 split roles; Q4 object-group leakage; Q5 changed measurements; Q6 evaluation limits; Q7 model versus control authority. Every option has a diagnostic explanation. The lab includes label quality and asks for prediction, numerical evidence, scope and a counterexample.