Manipulation of Heavy Object with Rapid Force Fluctuation by Humanoids Based on Motion Failure Detection and Repeated Trials

Humanoid robots can manipulate heavy objects by using their body weight. However, in manipulating heavy objects such as wheelchairs, tall shelves, and carts, the robot itself or the object easily falls over. Thus, proper control of the CoG and the manipulation force against the object's rapid movement is required. The following issues must be addressed in order for such manipulation. A humanoid is highly vulnerable to rapid force change and easily falls over if it improperly shifts the CoG. Further, it is difficult to estimate the complex manipulation force trajectory in advance if the object's precise physical and geometric parameters are unknown. In this paper, we propose a method to progressively acquire an appropriate force trajectory by repeating object manipulations. The proposed method solves the abovementioned problems as follows. The robot avoids falling over by detecting the motion failure in advance according to the robot's stability and the state estimation of the manipulated object while executing the motion. It predicts the correct force trajectory based on the force measured in this trial. It finetunes the predicted trajectory online with the measured reaction force. We implemented these methods into a control system and verified their effectiveness by experiments on an actual robot, realizing the pushing-up motion of the front wheels of a wheelchair by repeated trials, as an example.