Programming humanoid robots to execute arbitrary tasks remains difficult for numerous reasons, including high kinematic redundancy, complex dynamics, and the problems involved with balancing and locomotion. Additionally, "porting" a behavior to a robot with even slightly different kinematics or dynamics is typically infeasible. Behaviors are consequently often written anew, missing one of the key tenets of software development, that of component reuse.
In collaboration with Honda Research Institute, USA, we developed a framework, The Task Matrix, to mitigate the problem of skill transfer. The Task Matrix framework focuses on separating robot behaviors into robot independent and robot dependent components, the latter of which are minimal and standardized. The Task Matrix (from the definition a surrounding medium or structure) stores a set of robot-independent task programs, both primitive and complex (i.e., composed of primitives), Boolean conditions used to determine whether a task behavior can begin or continue execution, and motions. The Task Matrix is currently being used to develop behaviors for Honda's Asimo robot and for physically simulated humanoids.
We have implemented multiple, primitive behaviorsfor performing tasks in a robot-independent manner using the Task Matrix framework. These tasks include reaching to an object, positioning a tool, grasping and releasing objects, fixating on an object or another humanoid, mapping and modeling an environment, achieving a desired posture, and performing rote movements. These primitive behaviors are inspired from a work-measurement system (MTM-1) that is proven at decomposing complex occupational tasks into a set of primitive elements. Correspondingly, we have demonstrated that these primitive task programs can be combined to produce more complex behavior.
Honda Research Institute, USA
Evan Drumwright drumwrig@(remove)usc.edu