Neural Modeling of the Kinematics of Robot

We aim at building a joint space trajectory generation system. Connected to a xed manipulator with sensory feedback, neural networks are expected to move the end-eeector from any start to any goal connguration without colliding with obstacles. The output of our system is a series of consecutive conngurations yielding a joint-space trajectory. Sensory information serves as the only source of knowledge about the manipulator's kinematics and the trajectory features. This knowledge will gradually be incorporated into the trajectory retrieval system. The particular path planning problem we are currently tackling can be stated as follows: A 6-DOF manipulator with rotary joints, enclosed in a narrow cell is supposed to grasp an object in the cell. We do not presume xed obstacles, but allow slow changes of the environment. The manipulator has to move its end-eeector from a given current connguration to a goal position and orientation, stated in task space coordinates. In case there is a trajectory from the initial position to the target position stored in the \trajectory retrieval system", this trajectory will be returned stepwise to the manipulator (and possibly to the user). Exploiting the generalization abilities of a neural trajectory storage, initial and target position must not be exactly the positions originally used for trajectory generation. If there is no such trajectory stored, or if the returned trajectory can not be used any more due to changes of the environment, a new trajectory must be planned. Of course a new planned trajectory can be stored as well. Path planning poses two problems: The solution of the Inverse Kinematics problem to get the corresponding joint space positions and orientations, and the search for a joint space trajectory. Currently we concentrate on the latter problem and require the trajectory to meet the following conditions: the robot must not collide with the robot cell the robot must not collide with obstacles inside the cell none of the robot's limbs may bump into another limb the robot must comply with its joint limits some constraint function (usually the energy used for the motion) should be optimized. Sensory information, produced while the manipulator is moving, is the only source of knowledge about potential collisions. To our knowledge there is no fully automated solution to this problem. At the DLR 1 , for instance, the task is currently accomplished by manually controlling the manipulator's movement via a control