# Difference between revisions of "VVV09 Kinematic, my friend"

From Wiki for iCub and Friends

Line 1: | Line 1: | ||

− | VVV09 Kinematic, my friend | + | == VVV09 Kinematic, my friend == |

The aim subject is the learning of the forward kinematic model. | The aim subject is the learning of the forward kinematic model. | ||

Why learn the model, we already have an analytic method ? | Why learn the model, we already have an analytic method ? | ||

− | + | * A human like approach. | |

− | + | * To validate the feasibility of such kind of learning. | |

Why the forward model ? | Why the forward model ? | ||

− | + | * To keep all the information of the model (i.e infinity of pseudo-inverse of the Jacobian, null space computation). | |

Why the kinematic ? | Why the kinematic ? | ||

− | + | * Because the geometric model is not easily invertible. | |

What kind of algorithm is using ? | What kind of algorithm is using ? | ||

− | + | * LWPR, (Locally Weighted Projection Regression) [Vijayakumar 2005]. A powerfull online learning algorithm based on multiple linear regression computation in the space the most adapted (PLS). | |

What are the results ? | What are the results ? | ||

− | + | * The model is learned and uses "successfully" on the robot. | |

What about the "keep all the information" ? | What about the "keep all the information" ? | ||

− | + | * With the learned model we are able to control two end-effectors with a priority on one of them. | |

So, what did you do here ? | So, what did you do here ? | ||

− | + | * I tried to learn the model ... without the model but with the vision. | |

And ? | And ? | ||

− | + | * I have learned the model by attended motor babbling (As we play with the "end effector" of a baby). | |

− | + | * I "wanted" to be able to track two objects (my end effector and another object ... But it is apparently not so easy :p) in order to complete the learning of the model still as if we were playing with a child. |

## Revision as of 10:49, 29 July 2009

## VVV09 Kinematic, my friend

The aim subject is the learning of the forward kinematic model.

Why learn the model, we already have an analytic method ?

- A human like approach.
- To validate the feasibility of such kind of learning.

Why the forward model ?

- To keep all the information of the model (i.e infinity of pseudo-inverse of the Jacobian, null space computation).

Why the kinematic ?

- Because the geometric model is not easily invertible.

What kind of algorithm is using ?

- LWPR, (Locally Weighted Projection Regression) [Vijayakumar 2005]. A powerfull online learning algorithm based on multiple linear regression computation in the space the most adapted (PLS).

What are the results ?

- The model is learned and uses "successfully" on the robot.

What about the "keep all the information" ?

- With the learned model we are able to control two end-effectors with a priority on one of them.

So, what did you do here ?

- I tried to learn the model ... without the model but with the vision.

And ?

- I have learned the model by attended motor babbling (As we play with the "end effector" of a baby).
- I "wanted" to be able to track two objects (my end effector and another object ... But it is apparently not so easy :p) in order to complete the learning of the model still as if we were playing with a child.