Learning task constraints in robot grasping

i) Task constraints can be used to bridge the gap between the high-level symblic reasoning and low-level sensory representations ii) Generative models are suitable for inferring regions of stable grasps. The framework can also handle the uncertainty in real world applications. iii) The idea is to integrate this work in goal-directed imitation learning [5]. iv) The challenge can be to collect enough training data for learning more complex BNs in real world applications.