Station 1 · Video 309
Same numbers: which chart tells the true story?
Two months of the shop's sales. Pick a question, then a chart type. Bit tells you whether that chart helps or misleads.
1. The boss asks
2. You draw it as
Under the hood: the chart rules
Compare groups: bars. Over time: a line. How the mix changes: stacked bars by week. Avoid pies with many slices or two dates to compare; eyes are bad at comparing angles. A bar chart axis must start at 0.
The data is the simulated milk tea chain used across the channel's data series.
Station 2 · Video 346
Who gets a voucher? Drag the threshold, watch precision and recall trade places
The model scores the chance of leaving for 59 member customers it never saw. Each dot is a customer: red really left, blue stayed. Everyone right of the purple line gets a voucher.
Picking the threshold is a business call. Plug in your own numbers (assumptions):
Under the hood: precision and recall
Precision: of the customers who got a voucher, how many really left. Recall: of the customers who really left, how many got one. Lower the threshold and you catch more leavers (recall up) but waste more vouchers (precision down).
Probabilities come from the LogisticRegression model in the video's notebook (176 customers to learn, 59 to test). At 0.5 and 0.3 the numbers match the video exactly. Voucher cost and customer value are assumptions.
Station 3 · Videos 348, 42, 241
Grow the tree deeper: when does the model start memorizing?
A decision tree asks yes/no questions to guess whether a customer will leave. Deeper trees ask more questions. The train score is on customers it has seen; the test score is on new ones.
Under the hood: why the test score drops
There are only 235 customers and 2 months of data. A deeper tree memorizes random details of single customers (some leaves hold just 1 customer), so it guesses wrong on new ones. Tip: always print the train score next to the test score; the bigger the gap, the more you should suspect overfitting.
Every score was really computed with DecisionTreeClassifier in the video’s notebook, depth 1 to 13 (13 is the unlimited tree).
Station 4 · Video 362
Keep the wheel at 10 rad/s when 4 cups land on the tray
The delivery robot measures its wheel speed 100 times a second and corrects the voltage. Turn the P, I and D knobs, then press Run. At second 2, staff put 4 cups on the tray.
Settings from the video
Under the hood: the motor model
This page runs the exact model from the video: a motor τ·dω/dt = K·u − ω − load (K = 2, τ = 0.15 s), voltage arriving 30 ms late, ±12 V limit, a 360-tick encoder averaged over 5 readings. The I term stops adding up when the voltage hits the limit.
Try the video’s challenge: raise D to 0.05 and see how much more the voltage shakes.
Station 5 · Video 363
Switch off ceiling markers and see if the robot still knows where it is
The robot drives 3 delivery laps. Counting wheel turns alone (odometry) drifts because the right wheel is worn. The EKF mixes in 8 ceiling markers to correct itself. Tap a marker to switch it off or on, then press Run.
Under the hood: what the EKF does each step
Every 50 ms the EKF predicts the new position from the encoders and its uncertainty circle grows. The upward camera looks 10 times a second; any marker within 1 m corrects the position and the circle shrinks. The more accurate source gets more weight.
This page re-runs the video's simulation on your device (right wheel 0.5% bigger, about 1% random slip per wheel). Slip is random, so each run differs a little from the video: odometry 2.9 m off, EKF 1.8 cm average error.
Station 6 · Video 364
Plan the delivery robot’s route with A*
This is the map the robot drew with LiDAR (light grey is unknown, treated as blocked). Obstacles are padded to fit the robot's body. Pick a table and watch A* find the way, then compare with Dijkstra.
Deliver to
Under the hood: how A* scores a cell
Each cell scores f = g + h: g is the distance already travelled, h is the estimated distance left to the table (8 directions, no corner cutting). A* always opens the lowest f, so it heads for the goal. Dijkstra drops h and spreads evenly in every direction: same shortest path, many more cells opened.
The map is a 5 cm grid, 160 x 100 cells, taken from the video's 43-second survey run. The video's challenge: move the pillar and compare the new route with the old one.
Found this useful? Send it to a friend
Everything runs on your device; nothing is sent anywhere. The shop, its customers and the robot are simulated for learning; the numbers come from each video's notebook and practice code.