You, the people, have the power, the power to create machines, the power to create happiness! You, the people, have the power to make this life free and beautiful, to make this life a wonderful adventure.
Artificial intelligence is having a revolutionary effect on industrial production. In the following article, Baidu Ventures Partner and Founder of Comet Labs Saman Farid shares his thoughts on recent changes and future trends in Machine learning-driven production philosophies and technologies.
Humanity’s biggest collective endeavor
The food you eat, the knick-knacks on your desk, the building materials your house was made from, the car you drive and the screen you’re looking at all were made somewhere — fruits of the labour of teams of designers, scientists, engineers, planners. Simply put, industrial production is one of humanity’s most impressive achievements.Since we started building things, we’ve striven to build them faster, more accurately and with less waste. We are now at the verge of extraordinary changes in our ability to create.
Imagination does not become great until human beings, given the courage and the strength, use it to create.
AI driven decision making, sensor networks, intelligent robotics and connected networks of autonomous systems are upending age-old notions of how production is done and bringing drastically more efficiency and flexibility to the supply chain. This translates into lower costs and a higher standard of living for everyone, with less waste, less pollution and less energy used. Workers can spend more of their time engaged in meaningful activities, teaching, learning, and caring for those around them.
To get a sense of the scale and importance of this change, it’s worth noting that in 2018 the vast majority of the US GDP was in the production of consumer goods, equipment, agriculture, construction and real-estate, all of which follow a similar pattern of production. In the developing world the reliance on “production” as an economic engine is even bigger, with collective annual production totaling hundreds of trillions of dollars and employing billions of people across traditional manufacturing, agricultural production and mining.
AI has is changing the way we produce 👷
A move from static to dynamic models for production processes
Traditionally, if you wanted something new built, an industrial engineer would spend time defining all the inputs to a process, specifying your desired output products, and quantifying all inputs to design a process. Through rigorous design and simulation, they would balance factors like throughput, acceptable tolerances, inventory on hand, defects and cost to arrive ata design that best fits the desired production process.(This same principle applies to a farmer deciding what to plant, how to lay out the seeds, grow and fertilize his crop; or a general contractor developing a building plan for a new skyrise.) In order to speed up production, engineers will often design machinery specific to a single process stepin a single factory, which makes repair and maintenance a headache, and breakdowns surprisingly common (and there can be hundreds of steps!).
These models are delicate creatures, and face challenges in implementation — things take longer than expected, raw materials show up too early or too late, too big or too small, or with an error rate outside of the acceptable norm. These deviations propagate through the system, and tend to leave the process operating outside of its “optimal zone” for at least part of the time.
In larger projects, industrial engineers spend the first weeks or months tweaking the production line to account for these differences. Not only does this create inefficiency in the process, but it also makes management reluctant to make large changes to the product. This is why every new model-year of a car is only cosmetically different, or has small tweaks.
AI is enabling these models to be designed, measured, and readjusted in real-time.
Tighter Feedback Loops
AI and advanced sensor technologies have allowed us to get a much more accurate, real-time and granular understandingof all the inputs into the system. When are people working? Of the 45,000 screws which arrived, which ones are slightly larger, which ones are slightly smaller? What is the price of this adhesive this morning?
We can also closely observe the outputs noting their quality and quantity levels at each step of the production process. But the best part is, we can also effect change; we can adjust our processes dynamically to go faster or slower, or to compensate for input deviations. For example, in a process where two parts are assembled, if some arrive slightly oversized, adjust the “partner” part to accommodate.
More flexible systems
Previously, robotics (and all industrial automation) were built on purely deterministic systems, which meant that any changes in position, shape and size would throw off the system. Each workcell and each robot was designed to accomplish a specific task.
As computer vision and sensor fusion have improved, so has the ability of robotics to dynamically adjust to changes in the environment. This has enabled robots to work more seamlessly with humans, allowing for redesign of the manufacturing line, and as a result, much higher utilization.
More efficient coordination
Large facilities previously had the advantage of economies of scale and efficiency, because they were able to better centralize resources, procurement, supplier negotiations etc. With AI, automated supplier selection and negotiation, real-time supply-chain financing and risk assessment and demand prediction are becoming more feasible and accurate, leading to a world where small production facilities are on a much more level playing ground compared to the larger players.
Changes in the design processes (upstream) and logistics (downstream)
This article is mainly concerned with the production process itself, however it’s worth noting that AI also enables innovations in thedesign process which precedes production, and in thewarehousing / distribution / logisticsof completed products and raw materials (some thoughts on opportunities in warehousing can be found in a previous Baidu post). Big leaps are happening in rapid iterative design and modeling, while robotics and real-time adaptive warehousing systems are transforming the way we think about inventory as well.
How is this so different from the way things used to be?
From the stone age to the industrial revolution
For most of human history, production was a cottage industry. Each family, town, or village produced goods based on the resources and skills they had using their own methods and systems.
The industrial revolution led to machines with far more physical strength and mechanical dexteritythan humans had, and this enabled production to scale dramatically.
A new system of production was born, and humans soon realized that standardization would have a big impact on productivity. After the laborious effort of setting up the massive machines and production lines, we wanted to use them as much as possible before changing the configuration, thus, the more we could produce the same product, the cheaper the unit costs would be. Standardization also had the added benefit of making parts interchangeable — if you ordered a certain type of screw, you knew what you were getting, and could rely on its consistency.
Subsystems full of 🔩 standardized components allowed us to quickly build more complex things like buildings, airplanes, ships, and cars. Even a simple product like an office chair — with its plastic injection-molded wheels, machined ball bearings, stitched fabric and hydraulic lift — is the result of the work of hundreds of engineers, designers, production managers and workers. The symphony of cooperation came at a cost, however: no single mind was any longer capable of understanding the design of all the components, the reason they were chosen, and the alternatives considered. In order to try to reach a more optimal production system, new tools were needed.
The machine that builds machines
In the 1890s Frederik Taylor was one of the first engineers to realize we needed a better system.We needed to design a machine that humans were a part of. The machine was part of a production “system” wherein we used math and statistics to make important decisions and humans and other machines to execute them. We had to decide things like:
- In what order should we place the steps in a production line?
- How do we match the rate of each process step so that there is no additional lag?
- How do we decide how much raw material (inventory) to keep on-hand to keep the production line fed?
- How do we adjust to unexpected changes in the quantity and quality of our supplied raw materials?
- How do we organize workers so that everyone is at the right place at the right time?
- How do we minimize waste and downtime, while maintaining consistent quality?
To address these questions, tools such as the Toyota Production System (TPS) and Six Sigma methodology and the academic field of Operations Research helped engineers, line managers and business people plan and implement their models.
Unfortunately, most of these models made one important assumption: relatively fixed inputs(raw materials, labor) and relatively fixed outputs. In other words, the production line couldn’t produce trash cans today and ceiling fans tomorrow, and the raw materials needed to arrive with consistent quality on a consistent schedule. We could model and account for uncertainty, but always at a cost.
On the messy reality of a factory floor, most principles espoused by traditional operations researchers turned out to be inapplicable. Even the most basic building blocks such as queuing theory couldn’t be relied on (or were too complex to use) in real factories. Instead, simpler tools were used by practitioners that relied on more heuristic measures like TPS’s pull methodology or the just-in-time principles.
We needed to design systems that could observe, and respond to changes in the inputs to the system.
What makes this so hard?
It takes years to design and perfect a production line, and it only works because of a critical component: human workers. Humans are amazingly versatile! We can, with minimal training, turn a screw, then open a drawer, take out sandpaper, and then glue together a series of parts.
Since humans are so versatile, most production lines start out with lots of manual labor. The plan is constantly changing, and since people can adapt so quickly, they are a much better choice than traditional robots. Little by little, once the process is formalized, automation is added. Automation then ensures that repeatable process steps are carried out by machines, designed to do the same task over and over, faster than a person and with unwavering precision.
This same principle applies regardless of whether we want to produce corn, office chairs, cars, T-shirts or oil tankers. Repetition is the best friend of traditional automation.
If you’ve ever asked a factory for a quote on a product, you’ll know their first question: “it depends on how many you want.”
Unfortunately, that isn’t how demand generally works. Despite (or because of) the best efforts of marketers, the items people want are constantly changing. The market is constantly dictating changes in price, and the reality is that we don’t know what the future holds (as shown in the graph below).
What compounds this problem is that the variability that we demand is also constantly increasing. We want different designs and colors, different shapes and sizes, more optional features on our cars, refrigerators, shoes, wallets and door locks. This is amazing as a consumer, but requires more and more flexibility from manufacturers (which generally means more inventory, more production lines, and higher costs).
As a result, millions of warehouses store unsold inventory, and innumerable spare parts sit unused “just in case” demand spikes. As a consumer, every item you buy subsidizes the cost of all the wasted production, wasted storage, and idle production lines and labor waiting for orders.It’s estimated that between 20%-40% of the cost of a product is from inefficiencies in the production process.
In order to avoid this, production that is generally small-batch is sent to “job shops” instead of traditional continuous-flow production lines. These have much more manual labor, are generally much slower, produce more waste and require more set-up time to build the same things.
The best way to predict your future is to create it.
— Abraham Lincoln
What lies ahead?
At Baidu Ventures we believe there will be a new generation of industrial giants born during the next few years who will have an impact on the world similar to what Henry Ford, Thomas Edison and Andrew Carnegie had in their age. This could take many forms, but some meaningful archetypes are below:
🛠 New equipment providers:Similar to GE, Siemens, ABB, John Deere, Boeing, Caterpillar, Honeywell, Mitsui, Schlumberger, we believe a new generation of startups will build new types of equipment and systems that will enable industry to be rebuilt.
🏭 New operators:Similar to United Airlines, Uber, Exxon Mobil, Rio Tinto, BaoSteel, Foxconn, PG&E, AECOM, P&G etc., we believe there will be a new generation of companies that use the new paradigm of AI-enabled systems to create new farm operators, factories, mines and transportation companies with completely new business models, cost-structures, and markets.
🌠 New contexts:We believe many geographies that industrialized later have the opportunity to leapfrog, and build businesses in new contexts that we’ve never seen before. Just like it was hard to see the smartphone and predict its impact on ride-sharing, we believe this new generation of intelligent decision systems will create new types of industries, and we hope to learn from entrepreneurs working in this space. (Email us if you have ideas!)
We’ve been fortunate to have already collaborated with entrepreneurs working on different aspects of this new world, such as:
Aqrose: Using computer vision to identify defects in manufacturing and make corrections in real-time, currently partnering with some of the largest manufacturers in the world.
Automation Hero: Automating internal business processes that used to be cumbersome and time consuming, improving businesses’ visibility into their operations.
AMP Robotics: Industrial sorting using computer vision and dynamically adjusting robot arms, starting with sorting of recycling.
Covariant.ai(formerly Embodied Intelligence): Highly intelligent and adaptive robotics, allowing real-time learning and adaptation to the environment.
Celi Engineering:Using advanced sensors and modeling to improve the efficiency of industrial fuel burners, starting with steel mills.
Cubeworks:Building smart-dust millimeter scale sensing and compute platforms, allowing computers to see and understand the world around them better than ever before.
Ripcord:Digitizing records automatically and efficiently, enabling rapid recall and better decision making.
Kebotix:Using AI to discover and create advanced chemicals and materials, with a full self-driving lab that tests and verifies materials autonomously.
Openspace: Passive data collection about construction projects to ensure accurate and timely completion.
Sensoro:Ultra-low power communications networks and base stations which enable sensors to collect and communicate data in every context.
Veo Robotics:Bringing together the flexibility and dexterity of human workers with the strength and precision of industrial robots.
If you’re reading this and thinking “I can (or already have) a better way to solve these problems” I’d love to talk to you. You can find me at s[at]bv.ai
I also welcome all feedback and corrections on the post, please reach out!
Many amazing people contributed feedback and assistance in the preparation of this article and I’d like to thank Adam Kell, Patrick Sobalvarro, Dale Rutstein, Seaon Shin, Naren Ramaswami, Julian Shapiro, and the Baidu Ventures team for their input.
About Saman Farid
Saman is a passionate supporter of early-stage startups. He has built three companies in the areas of E-commerce, IP television, and logistics management, and faced the challenges of building, focusing, and scaling a company. After two exits, Saman started Comet Labs – a fund & incubator focused on AI and Machine learning where he has invested in nearly 50 companies, including Airmap, Ripcord, Abundant Robotics, Cobalt Robotics, 3Scan, Saleshero, Otosense and more.
Saman now leads Baidu Ventures US investments, where, in additional to capital, he is leveraging Baidu’s considerable resources in datasets, technical talent, and domain expertise in an effort to turbocharge the companies he invests in.
Saman previously spent roughly 15 years in China, and has held positions at Honeywell, Verizon, Deloitte Consulting and Microsoft in roles ranging from R&D to operations optimization. He received a Bachelors of Engineering from The Cooper Union and an MBA from Tsinghua and MIT.