An information theoretical incremental approach to sensor-based motion planning for eye-in-hand systems

Our research is concerned with sensor-based motion planning (MP) for articulated robot arms with many degrees of freedom. The articulated robot arm (called robot from here on), equipped with a range sensor on the end-effector of the robot that gives the distances of the objects from the sensor, is required to plan and execute collision-free motions in an environment initially unknown to the robot.; In our view, sensor-based MP can be decomposed into two key sub-problems: (i) view planning, i.e. to determine the next sensing action, including which region to scan and from where to scan this region, and (ii) C-space expansion, i.e. to incrementally expand the C-space (configuration space) representation. These two sub-problems are essentially repeatedly solved in an interleaved fashion in the sensor-based MP.; We approach the view planning problem from an information theoretical point of view, introduce and develop a novel concept of C-space entropy which characterizes the uncertainty of unknown C-space. Sensing actions reduce the entropy of the C-space. In each iteration of the planning process, the next scanning action is determined by maximizing the entropy reduction, or equivalently, gain in information.; Because of the high dimensionality of the C-space of a typical robot manipulator, a, roadmap (Probabilistic Roadmap specifically) is adopted for C-space representation. We call our approach sensor-based incremental construction of probabilistic roadmap (SBIC-PRM). The crux of C-space expansion is to incrementally construct a roadmap as the sensor senses new free space in the physical environment. The evolving roadmap reflects the connectivity of known free C-space, $Cfree$, within which the robot carries out its motion to further sense the physical environment. The $Cfree$ and the roadmap expand as the sensor senses new free regions in the physical space. This process is repeated until certain conditions are met. For example, a final goal is reachable from (one of the landmarks in) the roadmap, i.e. a simple planner in the algorithm finds a collision-free path from one of the landmarks to the goal, or the goal is declared unreachable. We ensure that the new landmarks lie in the new free regions in the C-space, $DCfree$, that corresponds to the additional free regions in the physical space, $DPfree$ obtained in the scan. This guarantees that the landmarks in the roadmap remains uniformly distributed in the C-space.; Our Eye-in-Hand system is implemented on a PUMA 560 robot with a range scanner mounted on its wrist. We present extensive experimental results with our sensor-based planner running on this real test-bed. The robot is started in unknown and cluttered environments. Typically, the planner is able to reach (planning as it senses) the goal configurations in about 7–25 scans (depending on the scene complexity), while avoiding collisions with the obstacles throughout. The experimental results show the efficacy of the C-space entropy concept in exploring the environment efficiently, and the ability of the incremental roadmap construction to keep the roadmap from degenerating.