Towards Bringing Robots into Homes

Abstract:
Home robots need to navigate in a partially structure environment and detect objects on the floor and on tables. Navigation is considered solved using laser sensors, however object recognition requires other sensor modalities. The objective of this work is the merging cognitive robot tasks using one sensing modality: stereo vision. Obstacle avoidance and navigation use the detection of free space - the traversable space in front of the robot. Localisation in a home environment is solved by matching the free space to the map, which gives results similar to laser sensors but takes objects at all heights into account. Finally, the approach naturally leads to object detection: attention is placed around the free space region and structure recognition methods can be applied to their advantage. We will show results for all tasks in lab and home settings.