Rodney Brooks and Gary Marcus have written and spoken extensively about the current state of artificial intelligence and its limitations. Now the two are teaming up to bridge the gap between current AI and the robust AI of the future.

Brooks and Marcus are two co-founders of a new Palo Alto, Calif.-based startup aptly named Robust AI. Brooks will be the CTO, while Marcus is the CEO. Robust AI is trying to build an industrial-grade cognitive platform that brings common sense reasoning to robots. This is an extremely difficult problem to solve, they said. And while it doesn’t guarantee success, Robust AI has quite the team steering the ship. The other Robust AI co-founders joining Brooks and Marcus are:

• CSO – Mohamed R. Amer, Senior Technical Manager at the Center for Vision Technologies, SRI International
• COO – Anthony Jules, former CTO Formant, COO Redwood Robotics
• Henrik Christensen, Qualcomm Chair of Robotics and Professor at UCSD

Interestingly, Marcus, a New York University psychology and neural science professor and renowned cognitive scientist, sold his AI company Geometric Intelligence to Uber in December 2016. Marcus ran Uber’s AI lab for three months before stepping down.

But as The New York Times reported at the time of the deal, Uber was interested in Geometric Intelligence because it was “tackling artificial intelligence in a different way. Rather than taking just one approach like deep learning, Geometric Intelligence combines data scientists who use varying techniques to study artificial intelligence, including the Bayesian and ‘evolutionary’ methods.”

Christensen announced Robust AI during his keynote at the Robotics Summit & Expo, which is produced by The Robot Report and ran June 5-6 in Boston.

Building a cognitive platform for decision-making

There are a lot of robotics applications out there, but very few of them are actually intelligent. The robots are often brittle and difficult to program. Robust AI’s argument is that deep learning alone is enough to move the needle.

“What we are in the process of building is a general-purpose cognitive engine that can be used in many verticals — anywhere you encounter a dynamic, flexible world,” Marcus told The Robot Report. “There are no general-purpose tools to help robots make decisions in open-ended environments.”

Christensen told The Robot Report Robust AI wants to have a prototype of its platform ready within the next 12 months. The “AI layer” will be licensed to companies. Details about exclusivity haven’t been worked out yet, and Christensen said Robust AI is talking to a large set of companies to figure out what are the best verticals to get started in. He said markets that make sense include construction, last-mile delivery and manufacturing.

Christensen added that robots in the home, including elder-care robots, might also eventually come into play. “In the home, the margin is very small, so the home will probably come much later.”

To build its cognitive platform, Marcus said Robust AI will take a hybrid approach by combining multiple techniques, including deep learning and symbolic AI, which was the dominant paradigm of AI research from the mid-1950s until the late 1980s. This will, of course, be a controversial approach in AI and machine learning circles, as many will argue symbolic AI is antiquated.

“I’m known as a critic of deep learning, but that doesn’t mean we aren’t going to use deep learning,” said Marcus. “My critique hasn’t been that it’s not useful; my critique is that it’s just one tool among many. [Famed cognitive scientist] Geoffrey Hinton says we don’t need the old stuff because it’s old-fashioned like combustion engines. I argue there’s a lot of old techniques that we can meld with new deep learning techniques.”

Data, data, data

Christensen said Robust AI will start out using existing datasets for respective vertical markets, but it will eventually build its own datasets. The company is also pursuing strategic alliances with companies, including iRobot.

“iRobot has a lot of data about the layout of homes,” said Christensen. “So we’ve talked to Colin [Angle, iRobot CEO, co-founder and chairman] about possibly using some of their datasets.”

Brooks is co-founder, former CTO (1990 – 2008) and former board member (1990 – 2011) of iRobot. Brooks also co-founded Rethink Robotics, a collaborative robotics company that shut down in October 2018.

Christensen reiterated that Robust AI is focusing on the decision-making aspects of robotics. “There are lots of folks working on navigation and gripping,” he said. “We want to build an AI layer on top of this and build the reasoning capabilities that will make the robots smarter.”

“If you have a Roomba moving around a home and it’s in the bedroom, you should know what the neighboring rooms are. If you’re in the kitchen, there’s probably a refrigerator and a stove,” Christensen added. “We want to use this commonsense knowledge to make these systems more robust. If you’re running a deep learning algorithm and it’s in doubt about whether an object is a bed or a couch, the system will know that if you’re in the living room, it’s probably a couch. If you’re in the bedroom, it’s probably a bed.

How Robust AI started

While Marcus was writing his latest book about the limitations of AI, he realized robotics provided a great platform to try to solve some of the greatest challenges of cognitive science. And when he decided his next business was a robotics play, Marcus immediately turned to Brooks.

“Rod and I actually bonded over our skepticism over AI,” Marcus said. “We don’t think driverless cars are impossible, but people have underestimated the magnitude of the problem. We’re converging from two different points about why things have gone wrong in AI and what can be done to fix it.”

Marcus noted that there could be another “AI winter,” which is a term that refers to a period of reduced funding and interest in AI research.

“Everyone is excited about deep learning, but it’s been overhyped,” he said. “It’s not as powerful as it’s been made to believe. We can look at other approaches, which is what Robust AI is looking to do.”

Investment from Playground Global

Playground Global led a “substantial” seed round in Robust AI. A number of other undisclosed investors also participated in the round. Christensen said Robust AI has office space at Playground Global for the next 12 months.

“We have a very good deal with Playground Global,” he said. “So right now, the focus is on getting a prototype out. Initially we’ll have between 15 and 20 people to build the prototype.”

Marcus said other locations were considered for Robust AI, including Boston. But since most of its founders are on the West Coast and a office deal came about with Playground, Palo Alto made sense. “We haven’t turned out backs on Boston,” said Marcus. “Boston was certainly in the running. Maybe we’ll start an office there eventually.”

Christensen and Marcus both acknowledged that creating a commonsense cognitive platform for robotics is a tough task. But they’ve created an all-star team that’s up to the challenge and are hiring seven additional employees.

“We realized how frustrating state-of-the-art robots really are,” said Marcus. “We want to make robots much smarter by making them understand how the real world works. Machines don’t have to think exactly like humans, which have flexibility and reasoning. Humans have terrible memories, so we’d be foolish to make machines like that.”