Source: Company Press Release (
click here to go to the source)
Getting robots to grab objects correctly and quickly has seen massive improvements over the last few years. The technology has spawned several robotics companies now providing commercial applications in the piece-picking and e-commerce fulfillment space, including RightHand Robotics, Kindred.AI, and Osaro, among others. Another new company, which has its origins in the university research space, is preparing to "grab" its place in the robotics grasping arena -- Ambidextrous Robotics.
The Berkeley, Calif.-based company is developing proprietary robot grasping software that can allow robots to grasp almost any object. The company builds on patent-pending technology developed at UC Berkeley, specifically the Dexterity Network, aka Dex-Net, project, which uses simulation-to-reality transfer learning to automate the training of deep neural networks for various robots, grippers, and cameras. Many of the team members involved in Dex-Net, including Jeff Mahler, Ph.D., and UC Berkeley Professor Ken Goldberg, Ph.D., are now involved with Ambidextrous as the company's CEO and Chief Scientist, respectively.
Dr. Jeff Mahler and Dr. Ken Goldberg formed Ambidextrous Robotics, a startup that develops software for AI-based robotic picking in e-commerce applications.
The Dex-Net algorithms combine the simulation of thousands of 3D object models, analytical wrench mechanics, structured domain randomization, and synthetic point clouds to produce a deep neural network that can "efficiently compute robust robot pick points for novel objects without further training."
In January 2019, the group presented Dex-Net 4.0, which can now train policies for a parallel-jaw and a vacuum-based suction cup gripper on 5 million synthetic depth images, grasps, and rewards generated from heaps of 3D objects. With a physical robot with two grippers, the policy "consistently clears bins of up to 25 novel objects with reliability greater than 95% at a rate of more than 300 mean picks per hour."
Mahler, speaking with Robotics Business Review recently, said the team began considering commercializing the technology around 2017, after the Dex-Net 2.0 release. "We had been certainly thinking about commercial use cases for some time, because we had a lot of inbound interest from different industry groups, from manufacturing to logistics, to even folks trying to do stuff with service robots in the home," said Mahler. "So we kind of had the feeling that we were onto something, that industries across the space have been trying to do, but not been able to."
In 2018, the team was invited to give a demonstration of the technology to Amazon, and the enthusiastic and positive response by the e-commerce giant helped inspire the team even more.
Prof. Ken Goldberg is William S. Floyd Jr. Distinguished Chair in Engineering at UC Berkeley.
"[Amazon] was a part of the process of getting us really inspired to do this, because we demonstrated this to Jeff Bezos, and that's what got Jeff Mahler and the team excited about it," said Goldberg. "The team had a bunch of great builders, but it wasn't clear that they wanted to do the startup. But then after that experience [with Amazon], they were like, 'Let's do it.' "
The next step for the company was figuring out which parts of the technology they could commercialize upon. Because the Dex-Net project was a university project, several of the components were published during the course of the project. "Early on we published a lot of our code and a lot of the algorithms, and it's because we're academics. That's what we do," Goldberg said. "The question then becomes, well, what are you protecting?"
Goldberg said Ambidextrous Robotics has filed patents with some of the basic approach, and if it gets approved the company believes it will have some rights to protect that area. Second, the company has some trade secrets that didn't get published, which is based on building systems and putting all of the different components together. "Our system has been adopted by about 10 different groups, based on what we've put out on the web or publicly," Goldberg said. "But then we had a number of them come back to us and say, 'We want to tune this.' So that's a big aspect of what we've been working on, is rewriting the code and making it much faster. There's a lot of tricks in the trade to learn that."
Third, the company is working on extensions to make it faster and more robust, as well as the customers who are paying the company to develop versions of the system for them. For example, the robot arm they've been using was good for research students, but is too slow for industry applications.
Jeff Mahler, Ph.D., Ambidextrous Robotics
"We have a whole road map of how we can start speeding things up," said Goldberg. "One of them is to speed up the sensing, one is to speed up the computation and motion planning, and then the hardware arm itself. On all those fronts, we're making advances in how we can shrink the time down and increase the reliability. It's been the laser focus of the company -- how can we do this in a commercially viable system, making industrial quality software that will be hardened for real, ongoing use."
The company is working with some undisclosed clients to further develop the software and system, and Goldberg said the clients will share data with Ambidextrous, in order to compute the successes, failures, and fine tune parameters from the development.