How AI Can Make Smart Robots Smarter
There’s a term that’s been going around called “lights out factories” which refers to manufacturing facilities that are so automated that you don’t even need to leave the lights on because there are no humans around. A company called Rethink Robotics proposed that we turn the lights back on, and allow their robots to work alongside humans who would use a “monkey-see, monkey-do” approach to training these robots – or “cobots” as they’re often called. The robot would use computer vision to watch the human work and then mimic the workflow. Everyone got super excited and threw $149.5 million at the idea, investors like Goldman Sachs, Jeff Bezos, and General Electric. While it all sounded great on paper, Rethink Robotics closed their doors last month and investors have likely lost the majority of their money. So, what happened?
What Happened to Rethink Robotics?
An article by MIT last month talked about the failure of Rethink Robotics, a company we fawned over before because it looked so promising at the time. While robots are now doing an ever-growing number of tasks, they “are easily befuddled by real-world complexity and are mostly painful to program” says the article. Turns out that the robots just weren’t smart enough to emulate humans. Maybe it actually is a matter of working smarter, not harder. Now, we see a number of startups that are focusing more on the artificial intelligence aspects of robotics and less on the actual hardware. The problem is, many of them are in stealth and not saying much, which means it’s tough to find a list of all the startups out there trying to make robots smarter using artificial intelligence. Therefore, what we have here is not an exhaustive list by any means, but rather an example of seven startups trying not to become the next Rethink Robotics.
Using VR to Teach AI Algorithms
Formerly known as Embodied Intelligence, San Francisco startup Covariant.ai raised $7 million in seed funding last year to develop “AI software that makes it easy to teach robots new, complex skills.” The way they accomplish this is by having humans do actions in Virtual Reality (VR) and then having the robots learn from that data. The Chief Scientist of the company, Pieter Abbeel, is also the Director of the UC Berkeley Robot Learning Lab and three of his star students have joined him to build a solution that can easily teach robots how to perform advanced manufacturing tasks like assembling electronics or manipulating fabrics. While researching this interesting company, we also came across a competitor of theirs called Kindred.ai
Last year we published an article titled “6 Canadian Robotics Startups Not Called Clearpath“, and one of those was Kindred.ai which has since taken in a $28 million round of funding that closed just over two weeks ago with Chinese giant Tencent leading the round. This brings total funding for Kindred.ai to $44 million. Along with a new logo, they also have a new website which now talks about a real product that’s available. They’ve developed AutoGrasp™, an AI grasping technology which operates using the following four steps:
If during any of those steps the computer starts to have issues, they have “humans-in-the-loop” who jump in and show the robot what’s up. After a certain number of failures on the part of the robot are detected, the session is then passed to a human operator and the AI algorithm switches to learning mode. Again, an MIT Tech Review article does a good job of covering the startup and talks about Kindred.ai’s method: have robots observe how humans behave in a virtual reality setting and then learning the human’s behavior over time. For the techies amongst us, some of Kindred.ai’s papers are available on their technology if you want to dig into the details.
Smarter Robots as a Service
Founded in 2018, San Francisco startup Bright Machines raised some eyebrows with a $179 million Series A funding round that closed late last month. An article by Tech Crunch covered the event and talked about how the startup was spun out of Flex (FLEX), a massive electronics manufacturer with over $25 billion in revenues for 2018. A former co-CEO from Autodesk is leading the nearly 400 employees that already work there and he’s supported by senior executives from Google, Tesla, and Siemens. Bright Machines intends to develop a subscription service for the $12 trillion manufacturing sector so that manufacturing can be fully automated through machine learning. Since the solution uses computer vision, you’ll need to turn those lights back on. According to Reuters, Bright Machines has developed “micro-factories” made up of its software and robot cells which its customers are using in half a dozen countries to replace the people who assemble and inspect electronics.
Reinforcement Learning for AI
Founded in 2014, Berlin startup Micropsi Industries has taken in $9.5 million in funding so far to develop a trainable robot control system that can be used “off-the-shelf” with robots from companies like Universal Robots and ABB by simply plugging into the hardware. The solution is called MIRAI, and it’s a robot control system designed for complex assembly tasks. Says the company:
MIRAI-controlled robots perform movements that react to minute changes in the environment, as perceived by the robot with cameras or force sensors. MIRAI skills aren’t programmed, but trained – through demonstration, correction, and repetition. This allows users to create skills without writing a single line of code or modelling the problem.
This method of training is referred to as “reinforcement learning”, and according to some Chinese research firm nobody’s really heard of called QYREPORTS, there are a fair number of startups to be found in this space. One of them is Osaro, a company that offers “deep reinforcement learning technology.”
Founded in 2015, San Francisco startup Osaro has taken in $13.3 million in funding to develop their “proprietary deep reinforcement learning technology.” An article by the MIT Tech Review titled “This is how the robot uprising finally begins” talks about how Osaro trained a robot to put chicken pieces in bento boxes after a 5-second training demonstration. The article talks about how “an off-the-shelf camera combined with machine-learning software on a powerful computer nearby” can be used to train a robot how to perform tasks through trial and error, learning from past experiences just like a human child would.
Update 10/04/19: Osaro has raised $16 million in Series B funding to help invest in talent acquisition, international deployments, and advancing their product lines. This brings the company’s total funding to $29.3 million to date.
Autonomous Commercial Robots
Moving across the pond to London, England, we find MOV.ai, a startup developing an operating system for autonomous robots to function in commercial settings. Founded in 2016, the company has taken in $3 million in seed funding to automate smaller pieces of equipment found in commercial settings like pallet jacks, hospital cart movers and even mobile robots whose job is to snap pictures. In short, it reduces the complexity of deploying large numbers of robots in a commercial setting and expecting them all to behave themselves.
Spatial Artificial Intelligence for Robots and Drones
Staying across the pond we’ll find SLAMcore, another London startup that was founded in 2016 and has taken in $5 million in funding to develop “spatial artificial intelligence algorithms for robots and drones” and Toyota happens to be one of their investors. The acronym SLAM stands for Simultaneous Localisation And Mapping which is all about having robots and drones enter areas they are entirely unfamiliar with and then navigate whatever they find there while simultaneously learning more about the environment – you know, just how travelers acquaint themselves to a foreign country where the rules might be very different.
If those robots ever come to ‘Murica, they better learn to look both ways before crossing the street.
According to A.T. Kearney, “traditional manufacturing accounts for 16 percent of the world’s $80 trillion economy” which means there is a $12.8 trillion market that can be addressed here. This is all about software, not hardware. The challenge to make smart robots become smarter robots that can mimic humans can’t just be solved by throwing money at the problem. That’s evidenced by Rethink Robotics crashing and burning after burning through $150 million, while five of the seven startups we’ve talked about in this article have taken in less than $14 million in total funding. For those of you who are rightfully concerned that intelligent robots might be taking over your job soon, we just gave you seven more reasons to worry.
We think this AI-powered weight loss app could be a multi-billion dollar business - not because it's backed by the world's most sophisticated investors - but because it works. If you want to lose weight and keep it off for good, check out Noom. People who use Noom lose weight and keep it off for good.