Getting robots to do interesting things in the real world is tougher than you might think. Robot perception is noisy, sometimes unreliable, and execution is usually not the clean, deterministic thing we are used to in simulation. We’ve put in a bunch of work to show how we can robustly execute complex tasks in the real world, even when the robot is interfered with. Take a look at this video.

In the second half, we can see what happens when I grab the robot, shove it, move the object it’s trying to grab or undo all of the work it has been doing.

This is all done with our Robust Logical-Dynamical Systems, which are based on Riemannian Motion Policies as a way of reactive real-time control. Take a look at the paper on ArXiV to know more; a version of this work will also be at IROS 2019.