🤖 Module 5: Train Your Robot Arm (AI)

From demonstrations to AI control

Record training data, train an ACT (Action Chunking Transformer) model, and watch the robot perform the task on its own.

Learning Objectives

  • Understand training data: (observation, action) pairs
  • Learn that AI predicts "what comes next" from past frames
  • See how action chunking means predicting sequences, not single steps

Hands-On Activities (Robot)

  • Create a dataset: Record 20–30 episodes of pick-and-place (or chosen task). Ensure varied starting positions and clear movements. Partner monitors quality (camera in frame, no occlusions).
  • Train the ACT model: Use the training UI with adjustable hyperparameters (chunk size, epochs). Watch the loss curve. Discuss what "training" means.
  • Test the trained model: Run inference and watch the robot perform the task autonomously. Compare success rate across teams and discuss what made some models work better.

Off-Robot / Tactile Activity: ACT Prediction Game

🎯 "What Happens Next?"

Concept: ACT predicts future arm poses from past observations. This game builds that intuition without a robot.

Materials
  • Time-lapse sequence of the arm (5–10 frames from a pick-and-place or similar)
  • Printed frames or a slideshow
  • Paper + pencils, or 3 printed "guess" options per round
How to Run It
  1. Show frames 1, 2, 3. Ask: "Based on these 3 frames, what will the arm do next?"
  2. Kids draw the next pose or choose from 3 options (A, B, C).
  3. Reveal frame 4. Compare with predictions. Discuss what clues helped.
  4. Repeat: show frames 2, 3, 4 → predict frame 5. Then 3, 4, 5 → predict 6, etc.
Why It Works for Tactile Learners

They physically draw or select the "next" frame. The act of predicting—using only past information—mirrors what the AI does. No code, no robot, but the concept sticks.

💡 Pro tip

Use real frames from your robot's camera during a demo. Kids recognize "that's our arm!" and connect the game directly to the training they'll do in Module 5.

👥 Team activity

Teamwork

  • Pairs: one records episodes, one monitors quality (e.g., camera in frame, clear movements)
  • Group: each team trains on a different task, then demos to the camp
  • Reflection: discuss what made some models work better than others