Séminaire (organisé par l’équipe de recherche ASER)


Université de Toronto (Canada)

Are we there yet ? Robots That Learn from Experience, and Enhance Their Performance and Autonomy over Time

Mardi 8 octobre à 14h00 en salle C221

Résumé :

Until recently, robots have not been capable of understanding and coping with unstructured environments (like the ones humans work in) because their systems have relied on knowing in advance the specifics of every possible situation they might encounter. Consequently, if we want robots to work along-side humans in uncontrolled, (partially) unknown and changing environments, we must equip them with the ability to learn from new situations and adapt accordingly.

In this talk, she will present learning algorithms that enable robots to improve on a given task through repeated execution - like humans who acquire incredible skills through practice. These learning methods combine the best-of-the-two worlds, control theory and machine learning : controllers based on a-priori model information assure safe operation initially, while based on the data collected during operation the model can be refined and the robot ?s performance can be gradually improved.

She will present experimental results that show successful learning on different robot platforms :

  • a stereo-camera-equipped rover that learns to traverse rough terrain using vision only,
  • flying robots that learn to race in a slalom, perform aerobatics, and dance to music.

But are we there yet ? Do robots have the skills to learn on their own and operate in unknown environments ?

Questions such as : How can acquired knowledge be generalized and applied to new tasks ? How do teams of robots learn efficiently ? define new areas of research.

PDF - 3.7 Mo


FR SHIC 3272

Collegium UTC/CNRS