A family of models for manipulation planning

Manipulation tasks are those that cannot be accomplished without making and breaking contacts. A common requirement of planning algorithms is an ability to hypothesize control actions and then predict the evolution of the system using a mathematical model. This model must capture the physics relevant to the task being planned and do so at an appropriate level of details. In this paper, a collection of time-stepping models suitable for manipulation planning are proposed. The model with the highest fidelity is one that incorporates rigid body dynamics, joint constraints, and contacts with local compliance and Coulomb friction. The idea is to plan tasks starting with the simplest model, upgrading the model and replanning each time a plan is found, until a plan is obtained with the model of desired fidelity. This approach is illustrated through a two-finger grasp acquisition problem