tsitsimis / robot-handover

A controller for the humanoid robot NAO for interactive human-to-robot object handovers

Interactive co-manipulation agent

A handover algorithm based on the detection of colored objects in the scene and on deciding about the robot’s ability to execute stable grasping. Takes into account the relative positioning and orientation of the object inside the field of view of the robot's on-board camera.

Robotic platform

Humanoid robot NAO

• vision: monocular RGB camera
• manipulator: 5-DOF arms, 1-DOF hands

Detection

The object detection is based on simple color-based segmentation:

• RGB → HSV
• thresholding
• morphological filtering:
• select max-area contour

Classification

The robot extracts visual geometric features of the object (centroid, width, height, orientation, etc) and based on training examples of good, average and bad handovers decides if it can grasp it. The classifier used is a Support Vector Machine with gaussian radial basis kernel.

Visual servoing and end-effector adaptation

The robot's head follows the object by adjusting its yaw and pitch angles through a simple proportional control. Given that the training samples were captures in a fixed head orientation the hand's pose must adapt to compensate the camera frame-hand frame change.

Dependencies

• numpy
• scikit-learn
• opencv
• naoqi