Interactive Perception of Unknown Articulated Objects
Autonomous manipulation in unstructured environments requires the following perceptual capabilities:
- Object Segmentation
- Object Tracking
- Kinematic Modeling
In my research, I developed the above perceptual capabilities for
modeling planar articulated objects and general
3D articulated objects.
In addition, I proposed a relational reinforcement learning framework for gathering, generalizing and transferring manipulation expertise.
This algorithm is computationally efficient, handles occlusion, and
depends on little object motion.
I conducted experiments with everyday objects on a robotic manipulation platform equipped with either a camera or an RGB-D sensor. The results demonstrate the robustness of the proposed method to lighting conditions, object appearance, size, structure, and configuration.
This work represents a paradigm shift in the structure of robotic
It breaks the traditional boundaries between action and perception: deliberate interactions are used to reveal information that would otherwise remain hidden or difficult to interpret.
|Modeling 3D Articulated Objects|
|Modeling Planar Articulated Objects|
|Relational Reinforcement Learning of Manipulation Expertise|
|Segmenting 3D Articulated Objects|