We have succeeded in making HRP-3 grasp the objects that the robot sees.
This video shows the humanoid robot completing a chain of tasks such as bipedal walking, visual recognition, and multiple finger grasp.
Four fingers with 17 joints (13 active joints and 4 linked joints) are attached to the right hand of HRP-3.
The position and posture of the object that HRP-3 sees are estimated based on edge direction histogram matching between a picture image captured by a single camera mounted in HRP-3 and 3D model data of the object.
For the first step of multiple finger grasp planning, we use rectangular parallelepiped models for both the object and the grasping region of hand to determine the reference grasp posture. The final grasping posture can be obtained by testing the force closure for the grasping postures close to its reference.