Many of the tasks we wish to have HERB perform require finding complex, multi-step manipulation plans. I have been working alongside several colleagues in the Personal Robotics Lab to develop a task planning pipeline that HERB can use to solve difficult tasks.  The pipeline contains a perception system that allows HERB to fuze multiple sensor inputs and object recognition strategies to form a full 3D model of the environment.  The planning component of the pipeline employs multi-step planning algorithms to decompone a task into individual components quickly and reliably solved by sequential calls to multiple planners. We seperate post-processing from the motion planning to speed overall task time by only post-processing plans that will be executed.  Finally, we use perception to detect failure and recover by locally modifying an existing solution.  Most recently, we have demonstrated the use of this system on a table clearing task with HERB.