Camp curriculum · Project 1 of 5
🍬 Project 1: Candy Dispenser
Kid line: When the camera sees a face, the arm gives you a candy.
Who does what
| Younger (play) | Older (copy / modify) | |
|---|---|---|
| Skill | Preloaded one-bin dispenser. Face → candy from Bin A. | Copy that skill. Add Bin B and a boy / girl check. Then test how often the model is right. |
Learning objectives
- Perception — The robot only “sees” this camera frame. No face in view → no candy.
- Demonstration — The dispense motion is a kinematic replay you (or the instructor) recorded. The VLM does not invent the reach.
- Reasoning — If a face, then Dispense A. That is the recipe.
- Limits (older only) — Boy / girl is a guess from a picture. The model uses shortcuts (hair, clothes). It will be wrong sometimes. Wrong guess ≠ who you are.
What younger students should be able to say: “It gave candy because it saw a face. It waited when nobody was there.” What older students should be able to say: that, plus “It guessed boy or girl from the photo. A guess can be wrong. We checked.”
The finished skill (instructors build this once)
Name in Raven: Candy Dispenser What kids play: one bin, face → replay.
Hardware (per table)
- Raven arm + tablet
- Camera aimed at faces (tablet front camera or webcam on a stand). Not the board.
- Bin A of wrapped candy within the arm’s reach
- Bin B (older): second pile / bowl, different candy or same candy, clearly labeled
- Small drop spot (cup / plate) in front of the kid
Keep candies wrapped. One candy per successful trigger.
Motions to record (basic)
| Action | What to record |
|---|---|
| Dispense A | Reach Bin A, pick one candy, drop on the plate. Kinematic replay. |
| Stay | No motion. Nobody in the camera / no face. |
Copy this skill to every camp tablet.
When it fires (basic)
Use presence (face) if the tablet’s face detector is reliable.
Fallback vision prompt:
Do you see a person’s face clearly in the camera image? If yes, action Dispense A. If there is no face, the face is covered, or you are not sure, action Stay.
Do not ask name, age, or “are they happy?” in the basic skill.
Runner (what younger kids see)
- Start / watch
- Emergency Stop
- Live camera
- Candy arrives or the arm waits
No Studio, no prompts, no boy/girl.
Play (ages 7–10)
They only run the one-bin skill. They change who is in front of the camera.
- Nobody in frame. Arm should Stay.
- One face (Pilot or Spotter). Arm should Dispense A.
- Cover the camera. Stay. Uncover. Dispense again.
- Two kids in frame (optional): still one candy — “it saw a face,” not “it counted kids.”
- Swap Pilot / Spotter so both get a turn in front of the camera.
Got it: they can say that in their own words, and point at camera vs arm vs candy. Almost: got candy; you had to say why it waited when the lens was covered. Not yet: think the arm is broken when nobody is in view, or that it “knows their name.”
Do not run boy / girl on this track.
Copy / modify (ages 10+)
Copy first: duplicate Candy Dispenser.
Required modify: add a second dispense and a boy / girl choice.
Motions to add
| Action | What to record |
|---|---|
| Dispense B | Same pick-and-drop from Bin B. |
Vision prompt (boy / girl) — Prompt G
Look at the person in the camera. If you guess the person appears as a boy, action Dispense A. If you guess the person appears as a girl, action Dispense B. If there is no face, more than one face, or you are not sure, action Stay. This is a guess from one photo, not a fact about the person.
They may tighten the prompt (hair, clothing) and watch whether that helps or just makes it more confident and still wrong.
How they test the model (the point of the modify)
This is an accuracy lab, not a labeling ceremony.
House rules (say before the first boy/girl run)
- Volunteer only. Anyone can Stay out of the camera.
- The robot’s answer is “the model guessed A or B,” never “you are a boy/girl.”
- Laugh at the model, not the person. A miss is data.
- Prefer printed photos (magazine / printed faces) if anyone is uncomfortable. Same test, fewer feelings.
- Do not save photos. Do not type names into the prompt.
Test protocol
- Run on printed photo 1 (clearly presented as boy in the photo). Record: Dispense A / B / Stay.
- Run on printed photo 2 (clearly presented as girl). Record the action.
- Optional: two volunteers who are willing. Same log.
- Fool it: hat, hoodie, hair up/down, camera from the side. Log again.
- Count: how many agree with what the photo was meant to show? How many Stay? How many flip when you change clothes/hair?
What they should notice
- Face-present (basic skill) is easy. Boy/girl is hard and uses shortcuts — same pattern as greedy classification.
- Lighting, angle, and clothes change the guess.
- Stay (not sure) is a good model behavior, not a broken robot.
- The arm will still move on a wrong guess unless the Spotter Stops. They are responsible.
Done when: Bin B is recorded, Prompt G (or their edit) chooses A vs B, they have a tally of right / wrong / unsure, and they can say the older explanation.
Not required: more than two bins, names, “age,” cloud vs local settings, gripper validation.
Materials
Every table
- Arm + tablet with Candy Dispenser (one bin) loaded
- Bin A + wrapped candies
- Plate / cup for the drop
- Camera aimed at face height
Older tables also
- Bin B + candies
- 4–6 laminated face printouts (mix of presentations; not campers’ photos)
- Optional hats / hoodies for the “fool it” step
- Tally sheet: photo id | guessed A or B | stay | notes
If it fails
- Face too far / dark → step closer, more light. Stay is success when there is no face.
- Presence never fires → use the fallback “Do you see a face?” prompt.
- Boy/girl always the same bin → that’s a result. Log it. Try a printout. Don’t “fix the kid.”
- Wrong bin about to drop → Stop. Still a successful lesson.
- No second bin recorded yet → they can still play one-bin and only add B when the motion works.