A Generative Model Towards Conditioned Robotic Object Manipulation

In a collaborative scenario, robots performing com- municative and legible gestures would improve the safety and the naturalness of the interaction. In this study, we introduce a novel conditional Generative Adversarial Network (cGAN) for the specific problem of generating time-series related to human manipulation of objects with different characteristics. A two- steps process involves the generation of new data in a latent features space, then their decoding to the target domain through a pre-trained decoder. Our model allows the control over specific properties of the generated output. The long-term goal of our approach is to use the synthetic time-series to control the end- effector of a robot, to produce motions as communicative and implicitly informative on the object properties as humans ones.