Learning Transferable Push Manipulation Skills in Novel Contexts

This paper is concerned with learning transferable forward models for push manipulation that can be applying to novel contexts and how improve the quality of prediction when critical information available. We propose learn a parametric internal model interactions that, similar humans, enables robot predict outcome physical interaction even in contexts. Given desired action, humans are capable identify where place their finger on new object so produce predictable motion object. achieve same behaviour by factorising into two parts. First, we set local contact represent geometrical relations between pusher, object, environment. Then these contacts change throughout push. The our model. By adjusting shapes distributions over parameters, modify model's response. Uniform yield coarse estimates no available about context (i.e. unbiased predictor). A more accurate predictor learned specific environment/object pair (e.g. low friction/high mass), i.e. biased predictor. effectiveness approach shown simulated environment which Pioneer 3-DX needs provide proof concept real robot. train 2 objects (a cube cylinder) total 24,000 pushes various conditions, test 6 encompassing variety shapes, sizes, parameters 14,400 predicted outcomes. Our results show both predictors reliably predictions line outcomes carefully tuned physics simulator.