Detection of kinematic constraint from search motion of a robot using link weights of a neural network

In this paper, a method for detecting kinematic constraints in a plane when the shapes of the grasped object and the environment are not given is presented. This method utilizes the displacement and force information obtained by "active search motion" of a robot. A new neural network configuration for this detection is proposed. It consists of two multilayer networks (primary and secondary network). The primary network learns the movable space (constraint) obtained by the search motion. By the generated link weights which reflect the movable space, the secondary network determines the type and the orientation of the constraint. Simulation and experimental results are presented and analyzed.

Authors

• Hiroaki Seki
• Ken Sasaki
• Masaharu Takano

Keywords

• Motion detection
• Kinematics
• Robots
• Neural networks
• Orbital robotics
• Friction
• Force sensors
• Object detection
• Shape
• Humans
• Neural Network
• Network Weights
• Link Weights
• Kinematic Constraints
• Primary Network
• Secondary Network
• Root Mean Square Error
• Stiffness
• Mean Square Error
• Output Layer
• Hidden Layer
• Unit Vector
• Output Value
• Unit Of Activity
• Forgetting
• Reaction Force
• Maximum Displacement
• Search Direction
• Output Units
• Types Of Constraints
• Error Backpropagation
• Range Of Space
• Weight Pattern
• Orientation Vector
• Robotic Hand
• Training Data