Motivation

To this day the most successful applications of articulated robots can be found on factory floors. This means: in completely controlled environments. Every little detail is designed to suit the limitations of the robot and the methods making it move. Assembly cells are commonly surrounded by walls to prevent anything unforeseen from happening, like a human getting into the way of the robot. They have limited sensing capabilities and basically no way to react to an unplanned event (if they can detect it) other than stopping and waiting for a human to intervene.

The situation is vastly different if we start thinking about populated and changing environments. Like a construction site, for example. It is impossible to design the environment to suit the robot. The robot cannot be stationary, as in the the car factory, but has to move about. It has to be able to move robustly in spite of constant changes in the environment. Hence, suitable methods to generate the motion of the robot have to address mobility and incorporate ways of reacting to unforeseen changes. This provides the motivation for the Elastic Strip Framework.

Planning and Execution

Traditionally, the algorithmic determination of motion has followed two different paradigms: planning and execution. In planning a motion is determined based on information about the entire environment before it is performed. Execution methods (reactive methods, control methods), on the other hand determine motion command during the motion, usually based on measured state information, such as distance to obstacles or joint angles.

Planning methods are computationally complex, because they consider global information. Using that global information makes it possible to ensure that a particular goal can be achieved (completeness). On the other hand, the time required to compute a motion is generally too long to accommodate unforeseen changes in the environment.

Execution methods are very efficient, because they only need to consider local information. As a consequence of ignoring global information, though, a method might fail to achieve the desired result, even though it would be possible to achieve it if all information were considered. That is a big disadvantage, but on the other hand execution methods are well suited to generate complex, local behavior such as the walking pattern of a humanoid robot, or obstacle avoidance in a dynamic environment - something almost impossible to achieve with planning methods.

Integration

A successful approach to motion generation for robots in dynamic environments, it would seem, has to combine the properties of planning and execution methods. Consequently, it seems like a natural choice to attempt to integrate planning and control into a unified motion generation approach.

Elastic Strips

Elastic strips are a framework to generate motion for robots - mobile and stationary, possibly with many degrees of freedom (joints) - in dynamic environments. They allow to combine various aspects of motion, such as task execution, obstacle avoidance, posture control, and transitioning between various task and behaviors. This is accomplished with an efficient and powerful integration of planning and execution methods.

Fundamentals of the Approach

Motion planning methods determine a motion by performing a global search in the high-dimensional configuration space. This search is computationally complex, but it enables some form of completeness. Execution methods avoid this search by only considering local information. They are very efficient and solve some problems elegantly, but might fail to achieve a global task. The elastic strip combines the advantages of both approaches into a single approach to motion generation.

The costly search in high-dimensional configuration space is replaced with a directed exploration of the search space, guided by efficient local methods. A tunnel of free space around a trajectory, which was previously computed by a global motion planner, is used to implicitly represent a set of homotopic paths (see figure). The tunnel effectively represent a set of alternative paths to the current one as a work space volume. The volume can be searched very efficiently for locally improved alternative plans using execution methods, such as potential field-based approaches or control methods.

This integration of planning and control via a set of implicitly represented homotopic paths permits the integration of global motion generated by a motion planner with advanced control methods to perform obstacle avoidance, task execution, posture control, walking pattern generation, balance control, redundancy resolution, behavior transitioning, etc. Hence, the elastic strip framework allows to address all aspects of a motion simultaneously, in real time, in dynamic environments, and for robots with many degrees of freedom. To see it all in action, check out the videos.