Lecture by Roberto Calandra (UC Berkeley)


Start date: 25/01/2018
Start time: 02:30 pm
End time: 03:30 pm
Location: SR 02-016/018, building 101, Faculty of Engineering

Title: Learning to Grasp from Vision and Touch

Abstract: Grasping and in-hand manipulation requires fine control of the contact forces. Accurately estimating such contact forces indirectly, such as from analytical model-based approaches or from vision alone, can be very hard. A simpler and more accurate estimate could be provided by the use of tactile sensors. However, processing tactile sensing for grasping is challenging, since analytic modeling of the contacts and its effect on tactile readings, as well as their relationship to the grasp outcome, is complex and highly dependent on an accurate geometric understanding of the scene. In this talk, we discuss how tactile readings can be integrated into a grasping system entirely through end-to-end learning from raw visuo-tactile data. Our approach is based on an action-conditional deep model that given visuo-tactile information about the current grasp and a candidate grasp adjustment, predict the success probability of the next grasp.This approach requires neither calibration of the tactile sensors, nor any analytical modeling of contact forces, thus significantly reducing the engineering effort required to obtain efficient grasping policies. We trained our visuo-tactile model with the data collected over 6,000 grasping trials on a two-finger gripper equipped with GelSight high-resolution tactile sensors on each finger.An extensive experimental validation show that our approach significantly improves a robot's ability to (i) correctly predict the outcomes of pre-recorded grasps, (ii) successfully grasp an object with the least number of attempts, and (iii) use minimal force while maintaining performance.

Bio: Roberto Calandra is a Postdoctoral Scholar at UC Berkeley in the Berkeley Artificial Intelligence Research Laboratory (BAIR) working with Sergey Levine. Previously, Roberto received his Ph.D. from TU Darmstadt (Germany) under the supervision of Jan Peters and Marc Deisenroth, an M.Sc. in Machine Learning and Data Mining from the Aalto University (Finland), and a B.Sc. in Computer Science from the Università degli studi di Palermo (Italy).His scientific interests focus at the conjunction of Machine Learning and Robotics, in what is know as Robot Learning. Some of the research topics that he is currently developing include Deep Reinforcement Learning, Bayesian Optimization, Dynamics Modeling, and Tactile Sensing.

Get this event: 
Subscribe to calendar: (Copy and paste into your host application - further infos here)