Computer Vision & Graphics Machine Learning & Data Science Research

AI Startup Building ‘Universal AI for Robots’

Covariant last month secured a US$40 million Series B funding round led by Index Ventures to push its total funding to US$67 million.

One could be excused for confusing the Investor’s list on the website with the all-star lineup for a top AI conference. The names include 2018 Turing Award winners Geoffrey Hinton and Yann LeCun, Google AI lead Jeff Dean, Director of the Stanford Artificial Intelligence Lab Fei-Fei Li and Berkeley Artificial Intelligence Research (BAIR) Lab Founding Co-Director Trevor Darrell.

Covariant last month secured a US$40 million Series B funding round led by Index Ventures to push its total funding to US$67 million. The ever-astute Professor Hinton recently tweeted he wishes he’d invested a hundred times more.

From academia to real world

Berkeley-based Covariant is building a universal AI designed to enable robots to see, reason, and act in the real world. “Covariant was founded with a very strong research DNA. And in a sense, you can say the company is on a quest to solve the hot research challenge of how do you build general AI for robotics,” Chief Executive and Co-founder Peter Chen told Synced. Initially known as Embodied Intelligence, the company was co-founded in 2017 by esteemed UC Berkeley professor Pieter Abbeel and colleagues Peter Chen, Rocky Duan and Tianhao Zhangfrom UC Berkeley and OpenAI.

The launch came with a mandate both ambiguous and ambitious, based on the team’s extensive expertise in artificial intelligence, deep learning, and robotics: “Traditional robot programming required substantial time and expertise,” explained Abbeel in a seed founding round press release. “What we will provide is an AI layer that can be added to any existing robot, enabling robots to learn new skills rather than requiring explicit programming.

Three years later, Covariant’s marque product is its Covariant Brain solution, which delivers “universal AI for robots that can be applied to any use case or customer environment. Covariant robots learn general abilities such as robust 3D perception, physical affordances of objects, few-shot learning and real-time motion planning, which enables them to quickly learn to manipulate objects without being told what to do.”

From advanced research techniques to robust applications in the warehouse, Covariant has developed and rolled out its solutions with impressive speed and success. In a 2019 global competition organized by Swiss-based ABB Robotics, 20 AI startups were tasked with designing software for ABB robot arms, with the efforts evaluated on 26 real-world picking, packing, and sorting tasks. Covariant won the global competition and entered into a partnership with ABB to co-develop AI solutions to assist autonomous materials handling. The first such deployment was at Active Ants, a leading provider of e-commerce services for web businesses in the Netherlands.

KNAPP, a world market leader in warehouse logistics and automation, has also been eyeing Covariant Brain. This March, KNAPP announced a partnership with Covariant to deploy and bring to market advanced AI Robotics solutions. Their first joint deployment was the Pick-It-Easy Robot at Obeta, a German electrical supply wholesaler located near Berlin. Jusuf Buzimkic, SVP of Engineering at KNAPP, enthusiastically endorsed the partnership: “We looked at every solution on the market, and Covariant was the clear winner. It can handle unlimited SKU types and works on challenging objects, including polybags, banded-apparel, transparent objects and blister packs. It also learns to pick new objects it’s never seen before and improves over time.” The New York Times reports the Pick-It-Easy Robot arm can handle more than 10,000 different items with better than 99 percent accuracy.

Tackling the tough part of logistics

How have Covariant-powered robot arms so successfully positioned themselves in terms of logistics and funding? The phrase “pragmatic research” surfaced repeatedly in our interview with CEO Chen.

“There are no textbook answers to a lot of things that we are trying to build. You need to embrace the fact that you need to solve unsolved problems. Then you need to assemble a team that could do that cutting edge research and advance over the state-of-the-art.

“Much amazing research talent is concentrated in places like Google, Facebook, DeepMind and OpenAI. They tend to focus much more on long term fundamental research. So this is typically a very long term problem that they’re aiming towards. What we try to do is be very pragmatic. Be very specific about the real-world problems that we want to solve, and get very deep into the problems, get our hands dirty in order to understand. You can say it’s a strategy, but it’s more of a culture, more of a mindset that we have to build up as a company.”

Chen told Synced that Covariant heard from the market that although many have approached picking problems in logistics and offered automated solutions, there as yet aren’t many large-scale AI systems available. He stresses that the new era of AI robotics is different from traditional industrial automation: “Now, your robot needs to deal with variability. It needs to deal with randomness. It needs to deal with changes in your environment.

It wasn’t a coincidence that Covariant chose to tackle the tougher part of logistics first. Abbeel told IEEE Spectrum that Covariant spent almost a yeartalking with literally hundreds of different companies [in electronics manufacturing, car manufacturing, textiles, bio labs, construction, farming, hotels, eldercare, etc.] about how smarter robots could potentially make a difference for them. Over time, it became clear to us that manufacturing and logistics are the two spaces where there’s most demand now, and logistics especially is just hurting really hard for more automation.”

According to Statista, retail e-commerce sales worldwide have reached US$3.53 trillion and are expected to grow to US$6.54 trillion by 2022. Accelerating global growth in the e-commerce sector has seen industrial leaders shift focus to AI-powered robotics solutions across a wide range of applications, including logistics, warehousing, and package sorting.

The Covariant Brain solution has so far provided use cases in various warehouse operations, where robotic arm applications mainly include depalletizing, picking, and sorting. Depalletizer systems unload and handle palletized products of different shapes and forms — such as boxes, trays, cases, sheets, bags, pails, and pallets. Warehouse picking meanwhile is where items are picked from a fulfillment facility to complete customer orders, and is one of the most expensive and labour-heavy processes in warehouses. Sorting is another essential warehouse operation, involving identifying items and sending them to the correct bin or storage area. Warehouse sorting robots are typically equipped with conveyors, arms, cameras and sensors; and rely on specialized algorithms.

Universal AI platform for robots

Covariant’s mission is to build a universal AI platform that enables AI robotics work autonomously in the real world. Rather than customizing entire AI systems to fit in different use cases in different environments, Chen says what’s important is a system’s ability to generalize to something new. “The world is constantly changing. Once you deploy a system into the real world, you need to face a lot of that.”

Covariant has built its business model to reflect its mission “A robotics system is not just the AI but also the robot itself, the surrounding, mechatronics equipment that you need to go with them. So the Covariant go-to-market model is what we call a development partner model. We will expose our Covariant Brain, this unified AI, as a platform, and our partners will develop on top of it.”

Chen says Covariant’s partnership with industry leaders such as Knapp and ABB also points to where a universal AI for robots might go next. “How do we as an AI software company help scaling these kind of applications quickly? This is not a well-known path. You could say there are obstacles, but maybe, in a sense, they are challenges that you necessarily need to face when you create a new category.”

Covariant plans to continue expanding into industries where robots are needed for repetitive tasks, such as food, healthcare, retail, parcels and manufacturing.

Journalist: Fangyu Cai | Editor: Michael Sarazen

6 comments on “AI Startup Building ‘Universal AI for Robots’

%d bloggers like this: