IPAB Workshop - 06/12/2018

Refining raw disparity maps from different algorithms to exploit their complementary advantages is still challenging. Uncertainty estimation and complex disparity relationship among pixels limit the accuracy and

robustness of the existing methods in the classical pipeline and there is no common method for depth fusion of different kind of data. In this talk, an architecture similar to GAN (generator is replaced by the refiner network without random noise input) is proposed to solve the three problems listed above, by designing an efficient network structure and an elegant robust object function. The fully supervised, semi-supervised, fully unsupervised methods have been explored to finish Monocular-stereo, stereo-stereo, stereo-ToF, stereo-Lidar depth  fusion. The proposed methods have been tested on the synthetic datasets (Scene Flow, SYNTH3, synthetic garden dataset ) and real datasets (Kitti2015, real Trimbot2020 Garden). Two depth-fusion application demos (autonomous driving and outdoors robot) in the real environment will be shown in this talk.

Title

Controlling and learning constrained movements for manipulation with contacts

Abstract

In contrast with many robotic applications where the environment poses an obstacle to the robot, in some tasks we require the robot to be in contact with it while executing a desired motion.

That interaction imposes a constraint to the robot motion and, therefore, we have to consider it when controlling robots or when learning those desired motions from demonstration.

In this talk I will be presenting two works where we addressed both controlling and learning a wiping motion with a robotic manipulator in contact with a curved surface.

The first work focuses on the robot control of a manipulator prototype for automating the cleaning process of a Train Cab front panels.

The second work, inspired by a similar wiping task, looks at how can we learn/unveil an underlying motion policy by several demonstration given data collected under different constraints.