We've seen incredible progress in robotics, however, getting robots to reliably manipulate objects, when there is error in sensing and action, is still a crucial challenge. Our goal is to improve the reliability of robot manipulation by creating models that predict the range of results from interactions between multiple objects and then using these models to plan for reliable manipulation. We learn the models from multiple simulations that model an unreliable robot pushing multiple objects. In these simulations we randomize characteristics such as the object mass and initial position to try to cover a wide range of outcomes. Then, we use methods from supervised learning to learn conservative models from the simulated data, that is, models that cover all of the observed outcomes and possibly more. We then use these models to plan a sequence of actions whose range of effects all achieve the goal, in particular, a desired object arrangement. We have shown that in some settings this approach is effective, both in simulation and on a real robot. We are working to extend the range of actions and object interactions that consider so as to enable construction of more complex object arrangements.
If you would like to contact us about our work, please scroll down to the people section and click on one of the group leads' people pages, where you can reach out to them directly.