This thesis presents a new robot learning framework, its application to exploit prior knowledge by encoding movement primitives in the form of a novel motion library, and the transfer of such knowledge to other robotic platforms in the form of shared latent spaces. In robot learning, it is often desirable to have robots that learn and acquire new skills rapidly. However, existing methods are specific to a certain task defined by the user, as well as time consuming to train. This includes for instance end-to-end models that can require a substantial amount of time to learn a certain skill. Such methods often start with no prior knowledge or little, and move slowly from erratic movements to the specific required motion. This is very different from how animals and humans learn motion skills. For instance, zebras in the African Savannah can learn to walk in few minutes just after being born. This suggests that some kind of prior knowledge is encoded into them. Leveraging this information may help improve and accelerate the learning and generation of new skills. These observations raise questions such as: how would this prior knowledge be represented? And how much would it help the learning process? Additionally, once learned, these models often do not transfer well to other robotic platforms requiring to teach to each other robot the same skills. This significantly increases the total training time and render the demonstration phase a tedious process. Would it be possible instead to exploit this prior knowledge to accelerate the learning process of new skills by transferring it to other robots? These are some of the questions that we are interested to investigate in this thesis. However, before examining these questions, a practical tool that allows one to easily test ideas in robot learning is needed. This tool would have to be easy-to-use, intuitive, generic, modular, and would need to let the user easily implement different ideas and compare different models/algorithms. Once implemented, we would then be able to focus on our original questions.
Exploiting Prior Knowledge in Robot Motion Skills Learning
DELHAISSE, BRIAN
2020-02-13
Abstract
This thesis presents a new robot learning framework, its application to exploit prior knowledge by encoding movement primitives in the form of a novel motion library, and the transfer of such knowledge to other robotic platforms in the form of shared latent spaces. In robot learning, it is often desirable to have robots that learn and acquire new skills rapidly. However, existing methods are specific to a certain task defined by the user, as well as time consuming to train. This includes for instance end-to-end models that can require a substantial amount of time to learn a certain skill. Such methods often start with no prior knowledge or little, and move slowly from erratic movements to the specific required motion. This is very different from how animals and humans learn motion skills. For instance, zebras in the African Savannah can learn to walk in few minutes just after being born. This suggests that some kind of prior knowledge is encoded into them. Leveraging this information may help improve and accelerate the learning and generation of new skills. These observations raise questions such as: how would this prior knowledge be represented? And how much would it help the learning process? Additionally, once learned, these models often do not transfer well to other robotic platforms requiring to teach to each other robot the same skills. This significantly increases the total training time and render the demonstration phase a tedious process. Would it be possible instead to exploit this prior knowledge to accelerate the learning process of new skills by transferring it to other robots? These are some of the questions that we are interested to investigate in this thesis. However, before examining these questions, a practical tool that allows one to easily test ideas in robot learning is needed. This tool would have to be easy-to-use, intuitive, generic, modular, and would need to let the user easily implement different ideas and compare different models/algorithms. Once implemented, we would then be able to focus on our original questions.File | Dimensione | Formato | |
---|---|---|---|
phdunige_4182849_1.pdf
accesso aperto
Descrizione: part 1
Tipologia:
Tesi di dottorato
Dimensione
12.28 MB
Formato
Adobe PDF
|
12.28 MB | Adobe PDF | Visualizza/Apri |
phdunige_4182849_2.pdf
accesso aperto
Descrizione: part 2
Tipologia:
Tesi di dottorato
Dimensione
18.47 MB
Formato
Adobe PDF
|
18.47 MB | Adobe PDF | Visualizza/Apri |
phdunige_4182849_3.pdf
accesso aperto
Descrizione: part 3
Tipologia:
Tesi di dottorato
Dimensione
8.07 MB
Formato
Adobe PDF
|
8.07 MB | Adobe PDF | Visualizza/Apri |
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.