News/Research

Ken Goldberg on AmbiRobotics on the Robot Report

04 Jun, 2021

Ken Goldberg on AmbiRobotics on the Robot Report

BCNM faculty member Ken Goldberg recently appeared on The Robot Report Podcast with colleague Jim Liefer of Ambi Robotics to discuss Ambi Robotics' recent US $6 million Seed round and its innovative picking robots. Ken describes the process in which his team and him developed Dex-Net, the operating system that these picking robots are based on. He also shares the lessons they have learned over the years such as the challenges of universal grasping and more.

From the podcast:

We've been working on this problem for 35 years - since I was an undergrad. We have been looking at the idea of how do we grasp objects and in particular objects of different shapes and sizes because it's fairly easy to grasp a cube or cylinder but once you get into more complicated shapes the problem is very difficult. It's actually challenging because humans are very good at this. Humans, even at a young age, babies can handle almost any shaped object and it will pick it up no problem, but robots really have a problem - they're still klutz. So we've been working on tackling this from a variety of perspectives. One of the big breakthroughs came when Jeff Moller, the cofounder of Ambi, joined our lab and we started working together to apply deep learning to the problem. We combined ideas from analytic models that go back almost a hundred years to be able to understand the physics and the mechanics of grasping with deep learning to be able to address the perception aspect of it. We also combined it with 3D depth sensing and then trained a very big neural network on lots of examples in simulation - that was a big differentiator because we had seen the success of this other system called ImageNet, which was trained on lots of labeled images, for learning how to recognize objects in a scene. Our idea was to apply the analogous idea to grasping.

Check out the entire podcast here!