Karlsruhe Institute of Technology (KIT), Germany - ARMAR-III

In this talk

1. 2:05

   Learning from prior experience

   The Xperience project explores how robots can use existing knowledge to learn more efficiently. Its structural bootstrapping approach combines prior experience with new observations to fill gaps about objects, actions, and planning.

2. 3:35

   Planning and setting a table

   In a dinner preparation example, ARMAR-III turns a spoken request into a planning goal, checks it against what the robot knows, and executes the resulting plan. For grasping, it selects a useful camera viewpoint based on uncertainty, then uses visual servoing to guide its hand.

3. 6:07

   Finding substitutes and locations

   When an object is unavailable, the robot can suggest an alternative using visual shape and affordance data, or commonsense relationships learned from text. It can also search remembered and typical object locations; when those fail, it asks a person for guidance and updates its memory.

4. 10:05

   Adapting learned actions

   ARMAR-III requests human help when an item is out of reach, and it can adapt a previously learned wiping motion to stir with a different hand orientation. It also adjusts wiping to the surface and current conditions, and uses a reachability map to position itself before moving a chair.

5. 13:28

   Human confirmation and task recovery

   The robot can offer a semantically related substitute, such as juice when soda is unavailable, and asks the person to confirm the choice. Its experience can also help locate an item in likely places and support replacing actions with similar ones.