Orbital AI Trials
I can label examples, train a decision tree, and test three unseen inputs.
Your activities & worksheet
Train a drone with labeled examples, then review its decisions in three new 3D test chambers. Work with a partner, swap roles, and record your own thinking.
What evidence shows that an AI can handle a new input?
Activity 1 · Train your drone
- Open Orbital AI Trials. Inspect training scenes in 3D.
- Label at least four varied scenes using at least two actions, then Train my AI. Explain a label before training.
Activity 2 · Review unseen predictions
- For each of the three test chambers, choose your predicted action.
- Review AI prediction and its learned path, then Test AI action in 3D. Record the independent task check.
Activity 3 · Improve and verify
- If a trial fails, review and correct a justified training label; retrain and test all three trials again.
- Complete the three unseen trials. Explain a learned split and why leaf label agreement is not a probability of correctness.
Individual exit check
Does perfect agreement on training examples prove that a new prediction is right?
Teach this lesson
A 50-minute plan with preparation, prompts, model responses, differentiation, and an individual assessment.
What evidence shows that an AI can handle a new input?
| Planning element | Teacher notes |
|---|---|
| Learning target | I can label examples, train a decision tree, and test three unseen inputs. |
| Evidence of learning | Train on at least four examples and pass all three unseen AI trials. Keep individual predictions, observations, and an explanation; a completion badge alone does not establish mastery. |
| Standards focus | Supporting investigation: 2-3.NSD.1; 4-6.CT.7, 4-6.CT.9 through explicit input/output tracing. AI is an enrichment context, not an additional claimed state standard. |
| Vocabulary | training example, label, decision tree, held-out input, prediction, test |
| Readiness | Follow a short sequence and use a grid. Revisit the previous week’s prediction and explanation before introducing this game. |
| Preparation | Open game-week-07.html (Orbital AI Trials). Check the four-rotor ai drone and Orbital Learning Station scene. Prepare a 5-by-5 grid, arrow counter, and worksheet. Use one device per pair and review the goal before revealing a solution. AI training labels and trial passes reset on reload; keep evidence in the worksheet. |
50-minute lesson sequence
| Minutes | Teaching moves | Look for |
|---|---|---|
| 0-5 | Introduce the mission Read the story and goal. Ask: What evidence shows that an AI can handle a new input? Identify what must be observed to verify success. | Learners restate the task and constraints. |
| 5-12 | Model one action or test Demonstrate one control without solving the whole game. Explain the difference between robot heading and camera view. Ask for a prediction before clicking Run. | Each learner names an input, expected action, and observable result. |
| 12-20 | Activity 1: Train your drone Open Orbital AI Trials. Inspect training scenes in 3D. Label at least four varied scenes using at least two actions, then Train my AI. Explain a label before training. Ask pairs to trace with a counter before retesting. | Collect the first worksheet prediction and observation. |
| 20-35 | Activity 2: Review unseen predictions For each of the three test chambers, choose your predicted action. Review AI prediction and its learned path, then Test AI action in 3D. Record the independent task check. Swap driver and navigator after one run. Pause at the first unexpected state. | Both partners connect a block, label, or parameter with the observed result. |
| 35-43 | Activity 3: Improve and verify If a trial fails, review and correct a justified training label; retrain and test all three trials again. Complete the three unseen trials. Explain a learned split and why leaf label agreement is not a probability of correctness. | A revision or comparison is supported by a recorded test. |
| 43-48 | Independent exit check Ask each learner: Does perfect agreement on training examples prove that a new prediction is right? Allow a spoken answer, drawing, or written response without a partner supplying the explanation. | Use the model answer as a guide; accept other explanations grounded in the run. |
| 48-50 | Save evidence and reflect Record completion and a next teaching step. Ask learners to download their worksheet or use a teacher-approved classroom record. | A next test or remaining limitation is named. |
Model program and expected behavior
Label four varied training scenes. Train. For each unseen trial: predict -> review AI -> test its action.
A useful four-scene set is 100→Move, 010→Right, 001→Left, 000→Stop (bits = ahead/right/left clear). A learned tree then passes held-out 110 and 101 with Move and 011 with an open-side turn. Trial inputs cannot be used as training scenes. The independent rule checks the prediction; it never replaces it. A label change clears the current trial passes and requires retraining.
Misconceptions and a useful response
| Misconception | Response |
|---|---|
| The answer check is the AI. | Compare the learned tree prediction with the separate rule check. Deliberately mislabel a training example and observe a wrong learned result. |
| Finishing the game proves I understand everything. | Ask for an individual prediction, trace, explanation, and a next test. Use completion as one piece of evidence. |
Differentiate the task
Support: Use path cards and one label at a time. Keep the explicit task rule visible. Allow spoken prediction and teacher-assisted selection. Provide the written position/heading status and a 2D or tactile map when the 3D view is not useful.
Extend: Add the optional 111 training scene or deliberately mislabel one scene, then explain changes in tree paths and test results. Identify the tiny dataset as a limitation.
Independent exit check
Prompt: Does perfect agreement on training examples prove that a new prediction is right?
Teacher key: No. Test new inputs and independently check the action. Training-label agreement is not a probability of correctness.
Connect to Scratch or physical robotics
A Scratch IF / ELSE is an explicitly authored rule. This local supervised model learns a rule from labels. Neither this activity nor the task check is an online conversational assistant.