Learned Models
Every kind of learned model a robot arm uses, in plain words
What a model is, how it learns and where its data comes from, for a reader who has never met one. Then classical machine learning, from regression and decision trees to Gaussian processes and movement primitives, and seven families of neural network models: models that see, that work in 3D, that choose a grasp, that move the arm, that understand words, that predict what happens next, and that make sense of touch and force. It ends with fine-tuning, running, testing and trusting a model on a real arm.
- 01
What Models Are
This is the first page of Learned Models, which explains the learned models that robots, and robot arms in particular, use today. Most of them are neural networks, although one chapter covers the older learning methods that are…
- 02
Classical Machine Learning
Not every learned model on a robot is a neural network, because the older learning methods, often called classical machine learning, are still used every day next to neural networks. This chapter covers those older methods, and…
- 2.1Classical machine learning14 min
- Most used2.2Linear and logistic regression22 min
- 2.3Decision trees and forests27 min
- 2.4Gaussian processes and Bayesian optimisation26 min
- 2.5Nearest neighbours and locally weighted regression22 min
- Also used2.6Mixture models and hidden Markov models21 min
- 2.7Movement primitives21 min
- 2.8Principal component analysis: shrinking data to a few numbers22 min
- 2.9Support vector machines19 min
- 03
Seeing Models
This page opens the chapter on seeing models, which are the models that turn a picture into names, boxes, outlines, poses or depth. They are the first of the seven kinds of model in this book, and a robot arm with a camera almost…
- 04
3D Models
This chapter is about neural network models that work in three dimensions. The last chapter, seeing models, worked on flat pictures, while the models in this chapter work on 3D points and whole scenes instead. So this page says…
- 05
Grasp Models
This chapter is about models that decide where and how to hold an object. A robot arm can see a mug on a table, and it can move its gripper to any point it can reach, but neither of those tells it where to put the fingers. So a…
- 06
Movement Models
This chapter is about models that move a robot arm. The earlier chapters of this book were about models that look at a picture and say what is in it, or where to hold an object. However, none of those models moves the arm itself…
- 6.1Movement models18 min
- Most used6.2Behaviour cloning49 min
- 6.3Action chunking transformers41 min
- 6.4Diffusion and flow policies42 min
- 6.5Actions and observations26 min
- Also used6.6Reinforcement learning policies39 min
- 6.7Learned motion planners45 min
- 6.8Reward and progress models41 min
- 6.9Learning from human video46 min
- 07
Language Models
This chapter is about models that work with words, and it answers three questions. What does a model that reads words do for a robot arm? How do words get into a model at all? And which of the three kinds of language model in…
- 08
World Models
This chapter is about models that predict what will happen next if the arm does something. This page is its overview, so it says what this family of models is for and names its four main kinds. It also shows how these models…
- 09
Touch and Body Models
This chapter is about the models that make sense of touch, force and the arm's own body. Every other chapter in this book starts from a camera, but this one starts from the sensors in the fingers, in the wrist and in the joints…
- 10
Making Models Work on an Arm
The seven chapters before this one each describe a family of models: what they take in, what they give back and how they are trained. This chapter is about what happens after that, because having a model is not the same as having…