Humanoid manipulation planning

	Topic	Paper page
	Push planning	Multi-modal planning for a humanoid manipulation task ASIMO Videos (user:anonymous, pass:isrr07)

Details

Description coming soon

Unifying multi-modal planning

In both locomotion and manipulation problems, the system moves between discrete modes. At each mode, the motion is constrained to a lower-dimensional submanifold of a higher-dimensional space. This mixed discrete/continuous structure is shared by a large number of systems (called hybrid systems in the controls literature). Some examples include:

• Legged robot locomotion, where each set of contacts (feet, hands, knees, etc) against the terrain yields different balance constraints. As above, a set of contacts is called a stance. Each stance corresponds to a mode.
• Object manipulation, where transit modes exist for each valid placement of the manipulated object, and transfer modes exist for each possible grasp of the object. In transit modes, the robot moves independently of the object, and in transfer modes, the object moves with the robot.
• Reconfigurable robots, where each discrete configuration is a mode.
• Assembly planning, where each set of disjoint subassemblies is a mode.
• Game-playing robots, where each step of the game rules yields a new mode.
• Other systems that make and break physical contact.

This work analyzes the properties of hybrid systems relevant to motion planning. Specifically, the regions of configuration space where the submanifolds of two modes intersect are of particular interest; they are precisely the regions in which the system can transition from one mode to another. Different planning approaches are appropriate depending on the dimensionality of these transition sets.

The paper above gives a randomized, general-purpose, multi-modal planning algorithm, which I believe can be shown to be probabilistically complete given certain assumptions. This work will be further explored in my thesis. The eventual goal of such work is to build a general-purpose multi-modal planner, whose only input is a model of the problem. Having such a planner could greatly reduce the effort needed to control high-dimensional hybrid systems.

Aside: I prefer using the term "multi-modal planning" when planning for hybrid systems. Occasionally this concept has been called "hybrid planning" from the field of hybrid control, but this term has been used (somewhat haphazardly) to refer to other concepts. It has also been called "multi-step planning" for legged robots, a term which is sometimes confusing (a motion for a person "taking a step" could transfer between three to five stances depending on the start and goal stance).