Autonomous Manipulation in Clutter
|Autonomous manipulation of unknown objects in a pile is
a prerequisite for a large variety of robotic applications
ranging from household robotics to flexible manufacturing
and from space exploration to search and rescue missions.
We focus on identifying and removing unknown objects from a pile, an important capability, enabling more advanced behaviors such as object counting, arranging, and sorting.
We propose an interactive perception approach in which the robot actively verifies its understanding of the pile. Our system is fully autonomous: the robot hypothesizes objects in the scene, verifies its hypotheses with deliberate interactions, and remove detected objects from the pile and into a target location. We rely on compliant motion primitives to guarantee the safety of the robot and the environment.
- The robot's perceptual capabilities include:
- A novel scene segmentation algorithm based on 3D data
- A novel object detection algorithm to verify segmentation hypotheses
- Autonomous action selection based on the verified segmentation
- The robot's manipulation and grasping capabilities are based on a library of compliant controllers
- Our system successfully integrates perception, manipulation, grasping, and motion planning