PacMan Leaflet v6 · Figure: A compositional representation of household objects. The objects share...
Transcript of PacMan Leaflet v6 · Figure: A compositional representation of household objects. The objects share...
What will we do?Imagine a robot performing an everyday manipula4on task such as loading a dishwasher. What does the robot need to know? How should it grasp a new object? Where should it place them in the dishwasher?
How does it gather the informa4on it needs to perform these tasks reliably? Pick up an object next to you now. Where did you look? How did you shape your hand? What informa4on did your fingers gather?
The PaCMan project will develop algorithms so that robots will be able to perform simple manipula4ons
on new objects reliably. PaCMan will focus on how the robot should internally represent the object’s proper4es that it learns through vision and touch. The representa4ons developed will enable a robot to
manipulate everyday objects, even if an object is new to the robot. The robot will autonomously gather informa4on about the object from vision and touch, plan its ac4vi4es, and check its ac4ons as it goes.
How will we do it?The PaCMan project is based on two core ideas: composi)onality and uncertainty. Composi2onality
means that our objects are made of a hierarchy of parts. The robot transfers informa4on between quite different objects that have similar parts. This part of the project will look at how to learn and use these
part based models of objects from vision and touch. Uncertainty is important for PaCMan in two ways. First the robot must know how uncertain it is about the object’s posi4on or shape. Second it needs to
know how this uncertainty will affect its planned ac4ons. This part of the project will reduce uncertainty by gathering informa4on.
Figure: A compositional representation of household objects. The objects share many parts. The simplest parts are just edges, the more complex parts include handles and long cylindrical parts
useful for grasping.
Proposed Work WP1 Learning Composi2onal Models from Vision will combine 2D and 3D visual informa4on
about object shape to learn a hierarchy of parts for recogni4on and reconstruc4on.
WP2 Learning Composi2onal Models from
Vision and Touch will build on WP1 by combining vision with touch. This will allow bePer learning
of 3D shape and surface fric4on.
WP3 Ac2ve Informa2on Gathering will develop techniques for ac4ve explora4on via touch and
vision to discover shape and fric4on. These will use part based models.
WP4 Grasping under Uncertainty will develop
new grasping methods, from hand pre-‐shaping to full grasping. We will grasp unfamiliar objects
using our composi4onal models and ac4ve touch.
WP5 Integra2on will bring together the different techniques and show them working on the
problem of loading a dishwasher.
Benefits and ImpactThe outcomes of this project will be more robust manipula4on of familiar and unfamiliar objects.
Extending manipula4on to real environments is one of the key scien4fic challenges leS in
robo4cs. This is required for most applica4ons of service robots.
PaCMan is funded by the European Commission
as FP7 project ICT-‐2011-‐9-‐600918.
Contact Informa2on
Professor Jeremy L WyaP
Intelligent Robo4cs Laboratory
Centre for Computa4onal Neuroscience & Cogni4ve Robo4cs,
School of Computer Science,
University of Birmingham,
Birmingham, B15 2TT, UK
www.pacman-‐project.eu