Camp curriculum · Project 2 of 5
🎯 Project 2: Intelligent Dispenser
Kid line: The robot still gives candy when it sees a face — but now it looks at the gripper. If it missed, it nudges and tries again.
Who does what
| Younger (play) | Older (copy / modify) | |
|---|---|---|
| Skill | Preloaded enhanced dispenser (check + retry). They compare it to Skill 1. | Copy Skill 1 (pure replay). Add pick-up check and left/right nudges. |
Skill 1 is the body (replay). Skill 2 is the same body plus intelligence (see if the pick worked, then recover).
Learning objectives
- Pure replay vs intelligence — Replay always finishes the motion, even with an empty gripper. The enhanced skill checks before it drops.
- Moving is not succeeding — The arm can go to the bin and still miss the candy.
- Perception (gripper) — A second look: “Is there candy in the gripper?”
- Recovery (older) — If the check fails, nudge Move Left or Move Right and pick again. That is a loop, not a new recording for every miss.
What younger students should be able to say: “The first robot just copied the move. This one checked. Empty gripper means try again, don’t pretend it worked.” What older students should be able to say: that, plus “We told the VLM to check pick-up. Nudge is a small fix, not a whole new dance.”
The finished skill (instructors build this once)
Name in Raven: Intelligent Dispenser Start from: Skill 1 Candy Dispenser (face → candy), but split dispense into Pick then Drop so a check can sit in the middle.
Younger kids play this preloaded skill. They do not build the loop.
Why split the motion
Skill 1 can be one recorded Dispense A (pick + drop). Skill 2 cannot: if pick and drop are one replay, the arm will drop even when the gripper is empty.
| Action | What to record |
|---|---|
| Pick A | Reach Bin A, close gripper on a candy, lift to a pose where the camera can see the gripper. Stop there. |
| Drop | From that lift pose, move to the plate, open gripper. |
| Stay | No motion (no face / give up). |
| Nudges | Do not record these. Use built-in Move Left and Move Right. |
Face trigger stays the same as Skill 1 (presence or “Do you see a face?” → start Pick A).
Pick-up check (Prompt V)
Run after Pick A, camera on the gripper (arm webcam if you have it; otherwise pause with the tablet aimed at the gripper).
Look at the orange gripper. Is a piece of wrapped candy clearly held in the gripper (not still in the bin, not on the table)? If yes, the check passes. If the gripper is empty, the candy was missed, or you are not sure, the check fails. Do not guess from the bin looking full. Only pass if candy is in the gripper.
Pass → Drop. Fail → corrective actions, then Pick A again (loop). Do not Drop.
Retry loop (instructor-built for the play skill)
On Pick A, enable vision validation:
- Prompt V
- Loop until success
- Corrective actions: Move Left, Move Right (and optionally Gripper Open then Pick A again)
Cap the loop (Raven max calls / instructor rule: 2–3 retries). After that, Stay and say “it gave up” — that is still a lesson.
The VLM should nudge toward the candy if it can see it in the bin, then the next Pick A replays.
Camera
- Face camera: same as Skill 1 (trigger).
- Gripper camera: needed for Prompt V. Instructors set this before class. If there is only one camera, the play skill still works if the lift pose puts the gripper in view of that camera.
Runner (younger kids)
- Start
- Emergency Stop
- Live camera
- They should see a pause after pick (“checking…”) then either drop or a small nudge and another pick
No Studio. No prompt editing.
Play (ages 7–10)
They already know Skill 1 (face → candy). Now they run Intelligent Dispenser and try to make it miss.
- Happy path: face in view, candy in easy reach. Pick → check → Drop. One candy.
- Replay contrast (if Skill 1 is still on the tablet): empty the easy candy / nudge the pile so Skill 1 would miss. Run Skill 1: arm still “dispenses” (empty drop).
- Intelligent miss: same awkward pile. Run Intelligent Dispenser: check should fail, nudge, retry. Candy or a clear give-up after a few tries.
- Cover the gripper camera during the check (optional): fail / Stay — “it can’t check if it can’t see.”
- Swap Pilot / Spotter.
Got it: can say replay vs check; empty gripper is a miss, not a successful candy. Almost: thought the nudge was “the robot dancing.” Not yet: think Intelligent Dispenser is broken because it paused, or that Skill 1 is “smarter” because it was faster.
Copy / modify (ages 10+)
Copy first: Skill 1 Candy Dispenser (pure replay), not a blank skill.
Required modify: turn it into Intelligent Dispenser.
- Split Dispense A into Pick A + Drop (re-record if it was one take).
- After Pick A, enable vision validation with Prompt V (they may rephrase; it must say candy in the gripper).
- Turn on loop until success. Add corrective Move Left and Move Right.
- Prove it: one easy pick (pass → drop) and one forced miss (candy offset) that nudges and retries.
- Optional: keep Skill 1’s boy/girl bins — only if Pick A / Pick B both get the same check. Do not require this.
Done when: empty gripper does not Drop; they can point at the check vs the recorded Pick; they can say the older explanation.
Not required: writing nudge trajectories by hand, boy/girl, extra bins, confidence numbers.
Materials
Same as Skill 1, plus:
- A way to force a miss (candy pushed to the side of the bin, or one candy barely in reach)
- Gripper in camera view at the lift pose
- Optional: Skill 1 still installed for the side-by-side contrast
If it fails
- Check always passes on empty gripper → Prompt V is too greedy (“I see candy” in the bin). Tighten: in the gripper only.
- Check always fails on a full gripper → light, closer camera, orange gripper in frame.
- Nudge makes it worse → smaller nudge / instructor demo table; still teach the idea with one retry Pick A and no nudge.
- Loop forever → cap retries; Stay is a valid ending.
- Don’t debug VLM download in front of kids.