AI is not just for driverless cars, digital assistants, or movie recommendations anymore; in multiple industries, it’s a wave about to break. According to a recent McKinsey Global Institute study surveying 3,000 “AI-aware” companies around the globe, only 20 percent are using AI-related technologies in a core part of the business, but the majority expect to ramp up AI spending in the next three years.
Other studies yield similar results. In an Infosys-funded survey of 1,600 business and IT leaders in seven countries, although only 25 percent said AI technologies were fully deployed and working within their organizations, 76 percent overall called AI critical to their companies’ success. Organizations with partially or fully deployed AI technologies expect them to contribute a 39 percent increase in revenue and a 37 percent reduction in costs by 2020. Businesses on average have been using AI for about two years and expect mature adoption in three more years.
Echoing insights from industry analysts and other observers, these reports conclude that AI in manufacturing is nearing a tipping point in the emerging factories of the future. There is consensus that AI applications ranging from smart and collaborative robotics to virtual assistants will upend how factories operate, requiring a complete rethinking of plant designs, manufacturing footprints, and supply chain models. So it’s not surprising that most AI tech investment has come from internal R&D dollars at big tech-savvy companies like Amazon, Baidu, and Google, according to the McKinsey study.
AI China 2020?
Most external investment is aimed at machine learning — the enabling technology for much of AI tech, such as speech recognition and robotics, according to the McKinsey study. Most of those dollars are going to U.S.-based companies (66 percent), followed by Chinese companies (17 percent), and both countries have established AI technology ecosystems.
The Infosys survey also found that U.S. organizations are the most likely to have increased their investments in AI technologies during the past year, while Chinese executives are the most likely to consider AI fundamental to the success of their business strategy.
Although China is a distant second to the United States in receiving AI investment dollars, its investment and deployment are growing fast. “Our survey suggests that countries that lead the way in AI investment and innovation — the United States and China — also lead the way in AI adoption,” the McKinsey study states.
Both countries
Infosys graph
In an Infosys-sponsored survey, adoption rates of AI technologies vary across vertical sectors, with manufacturing right in the middle, at 50 percent. (Source: Infosys)
have made AI technologies central to their strategies for the manufacturing sector. Last year, the Obama White House released a strategic plan for AI R&D, and China made AI a key part of its Five-Year Plan for 2020. AI is likewise critical to the success of Made in China 2025, the country’s new initiative to shift its manufacturing base from low-labor-cost to high-value-add manufacturing. To meet MIC 2025’s aggressive timetable, however, China may have to pursue more merger and acquisition deals along the lines of Chinese home appliance maker Midea’s 2016 acquisition of German-based Kuka Robotics.
The Infosys survey reports that China heads the AI maturity score by country, probably because it has “fewer legacy systems and business processes to contend with, making AI adoption and integration easier to accomplish.” But that doesn’t mean China will catch up anytime soon to the United States, which has advantages beyond its nearly fourfold lead in external AI investment dollars. For one, most of the biggest tech companies investing their own internal R&D dollars are U.S.-based, according to the McKinsey study. And the majority of the largest industrial manufacturers doing their own internal development — such as ABB, Bosch, GE, IBM, Siemens, and Tesla — are headquartered in the United States or Europe, not in China.
The United States has two major AI development hubs — Silicon Valley and New York — whereas China has only one, in Beijing, according to the McKinsey report. In electronics, multiple chip companies are developing or have announced GPUs or other processors that target AI; again, most of them are U.S.-based. When it comes to manufacturing itself, the renewed emphasis on Industry 4.0 advanced manufacturing [D3] is a strong indicator that the United States intends to reclaim at least some dominance on the factory floor. In turn, advanced manufacturing is one of the top four sectors in which the McKinsey study says AI will be adopted first, the others being financial services, retail, and health care.
AI in the U.S. factory
AI is being used in image recognition; security and financial transaction modeling; and process control in nonmanufacturing environments, such as oil and gas installations, utilities, customer service, financial services, and pharmaceutical supply. In factories, it’s handling such tasks as HVAC management and inventory management.
Outside the tech sector, AI technologies are still mostly experimental; with a few exceptions — notably automotive — very few factories have adopted them. Those that have implemented AI have started small, primarily in areas such as inventory management and inspection.
Machine learning is generally understood to include both training, as a machine learns to do something new, and inference, as it applies that learning to new information. For some, AI means the ability not only to create a model based on learning, but to change that model based on new data.
“True AI takes in data, creates a model, makes decisions based on that data, and then modifies that model,” said Jim McGregor, principal analyst for Tirias Research. “That ability is still limited, except for server-based solutions. Most companies that do manufacturing have some form of data analytics and automation control, which really isn’t considered AI but is rather a precursor to an AI system. Making this a true AI system would require training, monitoring, and modification/retraining based on information received.”
In most U.S. factories, there may be bits and pieces of AI in quality control and test but no complete systems, McGregor said. Neither models nor hardware systems have been perfected. Most AI applications today are either cloud-based or hybrids of AI running on the cloud and to the end node. “We're really just at the beginning of creating and optimizing systems that can be trained,” he said
New technology that will aid AI application development includes Nvidia’s Volta GPU architecture, the first to be built specifically for AI, said Murali Gopalakrishna, the company’s head of product management for intelligent machines. The first chip from this platform, the Tesla V100 GPU, delivers a 5x improvement over the current-generation Pascal, making it more efficient for machine learning, especially deep learning. Nvidia’s general-purpose GPUs have proved popular for training neural nets in machine learning to serve cloud AI applications.
For AI at the edge, Nvidia recently introduced the TX2 version of its Jetson system-on-module GPU-based computing platform. There are many use cases requiring AI at the edge, where lots of data must be processed in real time and decisions made immediately, instead of going back to the cloud first. “We’re seeing Jetson deployed on robots so they can be cognitive and smart and [can] influence their surroundings, although these haven’t been announced publicly,” said Gopalakrishna.
One concept starting to be used in optimization software is generative design, which uses AI to help optimize parts design, said Dayton Horvath, research associate for Lux Research. “AI can support operations like FEA [finite element analysis], used in modeling a simulation, and [can] handle tougher problems — ones with more degrees of freedom or with incomplete data sets,” Horvath said. For example, topology optimization can use AI to create a part that is lighter, yet with the same or greater strength, and heat exchangers can be modeled to be more efficient.
AI and robotics
One system mentioned often for factory AI applications is robotics. It’s already being applied in Universal Logic’s AI-based Neocortex technology, derived from NASA-sponsored research for the Robonaut space station robot. The technology lets automated systems handle deformable objects, high item variability, and parts changeovers without the need for fixturing.
Neocortex is the AI machine learning module of the company’s Spatial Vision 3-D software platform, which works with any actuated machine, not just robots, said Hob Wubbena, vice president of Universal Logic/Universal Robotics. The platform enables perception of the environment around the robot in applications such as machine tending and bin picking, and allows it to interact and respond to that environment in real time, at very high speeds. Capabilities include identifying and responding appropriately to a diverse mix of shapes and textures of objects, such as bottles, bags, and boxes, with 99 percent reliability.
For collaborative robots, human-in-the-loop reinforcement training will be key for making robots smarter with machine learning, said Erik Nieves, CEO and founder of PlusOne Robotics. “Reinforcement learning will have an impact both on the factory floor and on the distribution center,” he said. “Under the roof of every major factory is a distribution center, even if the people running the factory don’t think of it in those terms. But those who do are a lot further along in adopting AI for the factory.”
Two recent partnerships aim to develop AI technology specifically for robotics in industrial processes, including manufacturing. A suite of codeveloped solutions, including “cognitive industrial machines” that assist human workers in improving quality control, increasing speed and yield, and reducing downtime, will marry ABB Ability, a cloud-to-edge-to-device cross-industry digital capability, with IBM’s Watson Internet of Things platform.
The first codeveloped solution uses Watson’s AI and real-time production images captured through an ABB system to find defects, which are analyzed using Watson IoT for Manufacturing. IBM has commercialized the offering as the Cognitive Vision Inspection System, said Bret Greenstein, vice president of IBM Watson IoT.
Watson runs on the cloud, and subsets of it can run on servers. “We can even run it on edge devices and gateways, generally x86-based systems running Linux or embedded operating systems,” said Greenstein. “We are working with Cisco and others on this.”
In addition to supporting visual inspection, IBM uses Watson’s cognitive capabilities to interact with operators in hands-free environments or apply augmented reality to help diagnose and repair equipment. “We are seeing this adoption globally, in the U.S. and in other markets, as AI brings competitive advantages to improve quality, safety, and productivity, and enables building more sophisticated products,” said Greenstein.
Nvidia and Fanuc, meanwhile, are collaborating to add AI to the Fanuc Intelligent Edge Link and Drive (Field) system so robots in automated factories can work faster and more efficiently. The technology will use a range of Nvidia’s GPUs and deep-learning software to enable AI in the cloud, in the data center, and embedded within devices.
The Field system connects CNC machines, robots, peripheral devices, and sensors to optimize manufacturing production with analytics. Fanuc, which has recently demonstrated AI-powered robots, has three basic use cases for AI: pick and place, predictive maintenance at the edge, and automated optical inspection, for which inspection rates have already increased sevenfold, said Nvidia’s Gopalakrishna.
Nvidia Volta Tesla V100 Tensor cores
Nvidia says its Volta is the first GPU architecture to be built for AI, meaning training for machine learning. The Volta-based Tesla V100 GPU�s 640 Tensor cores deliver 120 Tflops � the equivalent performance of 100 CPUs for deep learning. (Source: Nvidia)
General Electric is developing its own technology internally to suit its specific manufacturing needs, as are some other vertically integrated U.S. manufacturers. In addition to building both hardware and software platforms, GE has invested in Clearpath Robotics, known for autonomous mobile robots, and OC Robotics, known for snake-arm robots, said John Lizzi, robotics technical operations leader for GE Global Research.
In some cases, GE is building robots from the ground up, such as units that can go inside of a jet engine to inspect it, and is building its own sensors as well as sourcing them commercially. AI becomes important in robotics via machine learning, which is a key technology for improving all three aspects of robots of the future: perception, advanced reasoning, and dexterity.
Collaborative robots are part of the bigger trend in robots, said Lizzi. The company’s vision is to move toward mobile, self-sustaining systems, with humans intervening only to handle exceptions, as well as to field smart robots that can work with humans in teams.
— Ann Thryft is the industrial control designline editor on EE Times.
Catch episodes 1, 3 and 4 in Made in China 2025 series:
- Part 1: Made in China 2025: Who Cares?
- Part 3: AI in US Factories? Not There Yet: July 27.
- Part 4: China's viewpoint: July 31.
Series editor is Junko Yoshida, EE Times. Global reports are written by journalists Sally Ward-Foxton, Ann Thryft, and Yorbe Zhang.
