Chapter 9 · Where to go next
Discover what the classical DMP cannot do yet, and find a path into the papers that explain each next step.
Your robot can now reach for a goal and keep a wiping rhythm, so you hand it a screwdriver and wait for it to turn its wrist to the right angle. It cannot do that yet. A chair moved into its path leaves it plowing straight on, and when a colleague grabs its arm mid-motion, it keeps pulling as though nothing happened. Those three encounters mark the edge of what you have built, and this chapter offers road signs toward the ideas that address them.
The picture
Imagine the moment after learning to ride a bike, when balancing, steering and braking begin to feel familiar and you look up at the roads ahead. Some climb hills, some carry traffic, and some lead toward cities whose names you have never heard.
The classical DMP is the bike you now know how to ride, and this chapter is a road map you can unfold beside it. You will find names and directions here, with a place to read about each route when you are ready to follow it.
The idea
You can carry the book’s central picture with you in two sentences: a spring pulls the robot to a goal, and a learned push shapes the path along the way. A clock tells the push when to act. As you explore the roads ahead, watch for the piece each one changes.
Intuition: one piece changes per road
Keep the spring, the push and the clock in view as you look over the map. Each road leaves two parts alone and changes the third, or adds a fourth part alongside them, so knowing which part changes gives you a foothold when you open a paper.
The signs below give each road a name and a place to begin reading when a question draws you that way.
Orientation. Return to the screwdriver and picture how the robot’s hand must turn: a position is a list of numbers that add up nicely, but two rotations do not just add. A DMP for that turning hand needs different tools, which you will find in S23 §2.1.2 for discrete motions and §2.2.2 for periodic ones.
Coupling terms. When the chair appears or someone pushes the arm, an extra push added to the DMP can help the robot react to the world. This is a coupling term, and you can read about using it to dodge obstacles or give way in IJ13 §3.2, then S23 §3.
Phase stopping, goal switching and other phase variables. So far the clock has run on its own, with the goal fixed during each motion. To explore what happens when you pause that clock, change the goal mid-motion, or choose a different kind of clock, follow S23 §2.1.1.2 and §2.1.1.3.
Generalizing to new situations. One demonstration gives you one motion, but imagine collecting several throws toward baskets in different places and wanting the robot to aim a throw toward a new basket. U10 labels each demonstration with its situation and uses regression to work out a new DMP for a situation it has never seen (U10 pp. 801–802).
Learning the frequency automatically. In chapter 7 you chose the periodic DMP’s frequency yourself; G09 explores how the robot can find it by watching a rhythm. Adaptive-frequency oscillators placed in front of the DMP discover that frequency and pass it along, leaving the DMP to learn the shape.
The survey itself. S23 has been your companion throughout this book, and it offers a place to begin exploring. Along the way, its list of open-source DMP implementations in §6.2 lets you see how other people bring the same ideas into code.
Further down the road, the survey points to probabilistic and neural variants, and to learning the weights by reinforcement learning. You can begin with its overview and follow the citations toward whichever question interests you.
When you are ready to set out, these four sources and their DOIs will help you find your way:
- S23, the tutorial survey (IJRR): 10.1177/02783649231201196
- IJ13, Ijspeert et al., Neural Computation: 10.1162/NECO_a_00393
- G09, Gams et al., periodic movements: 10.1007/s10514-009-9118-y
- U10, Ude et al., generalization: 10.1109/TRO.2010.2065430
The math
You have already gathered every equation you need to run a classical discrete or periodic DMP in chapters 1 to 8, so this chapter adds no new ones. Each road above brings its own equations, which are best read in the papers alongside the explanations that give them meaning.
Keep this book’s equations beside you as a dictionary when you read, because most extensions begin with the spring and the push you already know.
Try it
You can also keep exploring with your hands on the controls: the notebook below has both the discrete and the periodic DMP ready for something you imagine yourself. There is no separate sandbox notebook yet, so this is the place to start shaping a new motion.
Try it: 03_create_your_own.ipynb › Create 3.1 - Pick your letters
Choose your initials and watch two DMPs, one per direction, write them across the page; then let a different shape, goal or speed lead you into changing the code. See where you can take the robot beyond the motions this book has shown you.
Gotcha
Gotcha: papers use different conventions
You copy an equation from a paper into your code, and the robot does something strange, often because a convention changed along the way. This trips up everyone. Some papers scale the push \(f\) by \((g - y_0)\) while others leave it unscaled; some use \(\tau\) for the time scale, and some use \(\Omega\) for the frequency. Even a bump’s width may arrive as \(h_i\) in one paper and \(\sigma_i\) in another. The fix is to match every symbol against the equations of chapters 1 to 8 before you type a single line of code.