Method and manipulation system for manipulation of an object by a robot

A method for manipulation of an object by a robot, wherein the manipulation is directed at a manipulation target and comprises different manipulation modes, the method comprising:

• a) generating a plurality of manipulation mode-specific constraint-based controllers each for controlling a partial manipulation restricted to a manipulation mode;
• b) simulating, for a multiplicity of manipulation states and a multiplicity of the generated constraint-based controllers, in each case a partial manipulation which proceeds from the respective manipulation state and is controlled by the respective generated constraint-based controller, wherein an outlay value quantifying attainment of the manipulation target is determined;
• c) training, on a basis of the multiplicity of the manipulation states, the multiplicity of the generated constraint-based controllers and also the outlay values, a machine learning module to determine, on a basis of a current manipulation state, one of the generated constraint-based controllers which optimizes an outlay value;
• d) detecting the current manipulation state;
• e) determining an outlay-optimizing controller by the trained machine learning module on a basis of the current manipulation state, wherein the outlay-optimizing controller is configured to control a continuous and optimized path from the current manipulation state to the manipulation target;
• and
• f) generating control data for controlling the robot by the outlay-optimizing controller on the basis of the current manipulation state.