This paper describes the development of an autonomous serpentine robot. This robot is a 12 degree of freedom, untethered, hyper-redundant planar robot, named Kaa. Experiments were conducted in its application to mobility tasks that would be useful in locomotion among parallel pipe structures similar to those found in industrial plants. Kaa utilizes a behavior control system. Unlike classical approaches, that decompose the control system functionally, behavior control approaches decompose the control system by task. This control scheme requires very little computation; Kaa uses two 8-bit, 1 MIP microprocessors with a total of 36 Kbytes of memory. Only simple tactile sensing is employed; the distances between objects in the environment and the robot are neither implicitly modeled nor explicitly sensed by the robot. The behavior control system developed allowed this robot to reliably perform complex autonomous mobility and manipulation tasks with limited computation and sensing resources.

The authors focus on the crucial step in a pick-and-place operation, the choice of grasp. They describe an approach to choosing grasps, implemented in the task-level planning system HANDEY, which attempts to deal in a general fashion with the interaction between the choice of grasp and the choice of paths to reach the grasp. The approach is based on the systematic use of the configuration-space representation of the motion constraints on the robot. A very simple and efficient algorithm for approximating the configuration-space constraints leads to a practical method.>

We describe a robot system capable of locating a part in an unstructured pile of objects, choose a grasp on the part, plan a motion to reach the part safely, and plan a motion to place the part at a commanded position. The system requires as input a polyhedral world model including models of the part to be manipulated, the robot arm, and any other fixed objects in the environment. In addition, the system builds a depth map, using structured light, of the area where the part is to be found initially. Any other objects present in that area do not have to be modeled.