2 Jul 2018

Data, AI & Robots: Atomico’s Take on Industry 4.0

This post was co-authored by my fellow Industry 4.0 Team members Steve Crossan and Ben Blume, as we lift the lid on what is exciting us at Atomico at the moment about the future of industry

If you’d like more analysis and insight like this, delivered monthly(ish) to your inbox, sign up to The Operator’s Manual (our newsletter) here.

In a cable manufacturing plant near Chicago’s O’Hare International airport, a few small, sleek black boxes sit discreetly, plugged into decades-old plastic extrusion machinery, silently gathering data.

Distilled down locally on each of the boxes into a smaller set of meaningful variables, the data gathered is then wirelessly streamed to a cloud-based analytics platform, so that factory staff can monitor the production process in real time and from any device.

Oden Technologies — the company behind the platform (in which Atomico has just led a $10m Series A investment) — combines industrial hardware, wireless connectivity and a sophisticated data pipeline to produce an unprecedented (in this industry) view of the factory floor and its production processes.

At this deployment, issues are caught up to 95 percent faster (i.e. in minutes or hours, versus up to weeks), cutting waste by hundreds of thousands of dollars per year per plant, while increasing output by 10–15 percent through increased steady-state line speed.

Industry 4.0

Oden is one of a number of companies at the forefront of what is set to become an epochal shift in manufacturing (a $12tn sector globally, accounting for 17 percent of global GDP). The sector has so far remained relatively untouched by digital technology, but that’s changing fast.

Inexpensive sensors, cheap wireless communications infrastructure, highly scalable cloud-based data processing and novel machine learning methods have converged to a point where the building blocks are in place for a new Machine Age.

Dubbed Industry 4.0, these advances have not gone unnoticed by the world’s manufacturers. They have no choice: macro factors are forcing their hand. Fierce competition from nimble new challengers from China mean European and US manufacturers need to step up just to stay competitive.

A shift from mass, uniform manufacturing to small batch size, customized products — from personalized footwear to medicine — means traditional methods become unsuitably expensive. And customers, whether consumers or businesses, demand ever quicker turnaround times.

The holy grail for many is the “lights-out factory”: a factory that runs itself with zero human intervention (and presumably, then, no need for lighting). But how far away is that idea?

Tesla with its recent problems in Model 3 Production provides a cautionary tale. Though a visionary and laudable attempt, pushing automation too quickly and dogmatically can mean too many moving parts, chaos and confusion.

The more reasonable (but, in time, no less revolutionary) approach is to be incremental and iterative. By layering sensing and data gathering, new analytics, control systems, cheaper robotics and machine intelligence-based orchestration and analysis, and, by creating a culture of what in the software world is called “agile development” — i.e. moving quickly with imperfect information and biasing toward trying out new things — the more visionary factories in a good number of industries should be able to get to complete automation well within a decade.

Many in industry are not yet this optimistic. A recent PWC survey on Digital Factories showed that only 11 percent of companies that responded expected full automation within the next five years.

However, the silver lining was that two thirds expected to be using data for predictive maintenance, analytics, monitoring and optimization, and they estimated an average 12 percent gain in efficiency over the period.

Digital Transformation is Now a Given

At my own startup, The Climate Corporation, we found ourselves quite accidentally discovering a deep need for a large traditional sector (agriculture) to digitalize and to use data science to respond to major challenges, from climate change to feeding the appetites of a changing population.

The industry didn’t immediately embrace us, but the discipline we catalyzed, “Precision Agriculture”, is now broadly accepted. The very same principle applies in Industry too, which is why some of us at Atomico have taken such a keen interest here and count it as an area of ongoing focus.

Our friends at Point Nine Capital have written a great article (Reinventing the Factory Stack), listing over 200 startups working to disrupt this industry across 11 areas, from prototyping to workflow automation.

Venture capital investment in internet-of-things in Industry (the decidedly ugly-sounding “IIoT”) is at an all time high, according to a CB Insights report on the topic, with over $1bn invested in Q4 2017 alone.

Five Key Areas

We currently look at opportunities in the context of five areas which are converging to shape smart manufacturing: Analytics & Orchestration, Computer Vision, Robotics, AR/Wearables and Design/Prototyping. These are interrelated, but we still find this a useful segmentation.

  1. Analytics & Orchestration

It is said these days that “Data is the new Oil”. This is no less true in Industry as it is in other sectors. More data will be generated in the next two years than everything we’ve ever generated in human history — and this exponential growth will continue for a decade.

Obviously much of this data is raw sensor output that you can’t compare byte-for-byte in value with the works of Plato, but, nevertheless, we are going to be swimming in information. Thankfully, with machine learning we now have the means to make some sense of it.

Cheaper, powerful processors will allow much of this to be processed on-site (“Edge Computing”), avoiding an otherwise prohibitive bandwidth bottleneck.

We believe this automation will first mainly focus on measurement and alerting. Sensors will track metrics from machines that were never before measured during operation. Modern, scalable data platforms will connect data that was filed away into forgotten data silos.

Predictive algorithms and neural networks will work out relationships between the legion metrics to produce new insights into complex process chains. These will initially be advisory in nature, and humans make decisions on what to do when something is flagged to them.

In time (but before very long) control systems will be tied back to these models, allowing orchestration: software agents (“AI”) are given increasing agency to make decisions based on the observed data. They may even experiment with parameters to better model cause and effect and train themselves on the fly.

We’ve already seen this orchestration approach work resoundingly well in the context of data center energy optimization. Imagine the same happening in the manufacturing of medicine, with a machine overlord catching issues within milliseconds at fine granularity within complex process, automatically adjusting or discarding output early when unsalvageable — thereby avoiding highly expensive wasted batches, costly recalls, or, worse, adverse or fatal patient implications. This isn’t hypothetical — startup Bigfinite is already delivering this for some of their big pharma customers.

A recent BCG study reveals that a fairly high proportion of people in industry seem to get it: four in ten of the 1000 executives they surveyed expect AI to become “the key driver of productivity improvement by 2030” (most of the other 60 percent will no doubt change their tune within the coming few years).

2. Computer Vision

Over the last few years, computer vision is the sub-discipline of AI that has seen some of the most impressive improvements, surpassing human capability in image classification tests like ImageNet. Many tasks that previously required human visual inspection can now be done even better by machines.

Obviously, this has myriad applications within the factory setting — most obviously in quality control (QC) monitoring. Even the best human operators are inherently prone to inconsistency and unreliability due to distraction or fatigue.

Smart quality control startups like Scortex in France or Relimetrics in Germany deploy computer vision to automate visual inspection. This is one advancement that likely will come much sooner than later, given the limited dependencies involved in implementation.

Computer vision also forms a key component of the other trends listed here, particularly robotics and AR/VR.

3. Robotics and AVs

Robots have long been deployed in manufacturing to automate repetitive tasks, but these have been prohibitively expensive except for the largest scale use cases such as car manufacturing. These robots have also been highly inflexible, requiring precise calibration and will little tolerance for error.

Multiple trends, from small batch manufacturing, to interest in “re-onshoring” production in the face of rising labour costs and political/stability considerations has made small scale automation highly desirable — if it can be pulled off. Luckily, it increasingly can.

Small industrial robotic arms today are cheaper and more flexible than ever before, with the meteoric rise of smaller and more flexible collaborative robots — or cobots — set to irreversibly change the landscape. They cheaper, reprogrammable, and safer (so they can work alongside humans where previously they’d have to operate separately in cages — an important consideration for incremental introduction).

Several $25k robot arms coordinated by a single human can credibly perform tasks that would previously require 4–5 workers, at a cost savings of 3x or more.

Would-be customers are responding. Robot shipments grew 22 percent in the US in the first quarter of 2018. Meanwhile, the International Federation of Robotics, an industry body, estimates that the number of industrial robots in operation will double between 2014 and 2020.

We believe we’re at the beginning of a movement here, and the pace will very significantly increase. Exciting companies in this space include co-bot leader Universal Robots, Munich-based Kuka (now acquired by Chinese pioneer Midea), and challengers Franka Emika and earlier stage firm Automata.

Where this gets most interesting is in the software layer to teach robots to reliably carry out previously hard-to-automate tasks, including pick-and-place tasks, assembly, cable insertion, installation and device testing. A cluster of startups with very capable teams are leading the charge in this space, including Covariant.ai, Micropsi, Nomagic and Osaro.

Beyond robotic arms, other pieces of equipment within the factory can be automated — examples of this are forklifts or trolleys, enabled by the development of autonomous vehicles, which operate in simpler, less cluttered environments. Forklift & warehouse truck operators account for around $5.3bn of labour in the US alone, work that in large part is automatable. Globally this number is likely 3–4x the size. That’s one application.

4. AR/Wearables

In the UK alone, the annual cost of workplace injuries in the manufacturing sector averages around £500m. Factories are inherently rough-and-ready, physical places with people working alongside large, cumbersome and often dangerous machinery.

The potential for wearables and new materials to help here is significant, from AR/VR approaches to allow control from a safer distance to more out-there ideas like exoskeletons (see SuitX’s ‘space age’ industrial exoskeletons), or brain machine interfaces as developed by CTRL Labs to control robots far more naturally than old-fashioned joysticks and keyboards.

On AR/VR, a wide range of commodity hardware devices from AR glasses to smart watches and arm-mounted tablets can be woven together by vertically focused software companies to produce products that assist humans by overlaying relevant data into their field of view.

Zurich-based Scandit, an Atomico portfolio company, is doing interesting work in this area combining the humble but reliable barcode with modern computer vision methods to track objects through warehouses and supply chains.

This said, while AR is a good stepping stone and will remain valuable, the biggest driver of efficiency (and reduced injuries) will be putting fewer people into these risky jobs in the first place.

5. Tighter Design & Prototyping Cycles

Driven by huge improvements in recent decades in speed of fulfilment and delivery, expectations around the pace at which new products can be brought to market have rapidly increased, as well as demand for more customised and personalized products. Factories are being forced to adapt to support this to remain competitive.

3D printing has been a huge tailwind to the process or prototyping and designing, and it remains a powerful new tool in the design arsenal (particularly great in democratizing innovation in physical products).

This is now expanding into more traditional (and much lower cost, for a given quality) manufacturing methods such as CNC milling. Companies like CloudNC (an Atomico portfolio company) and Plethora are automating milling, allowing design engineers to identify manufacturing issues early and get back finished components within days instead of weeks to months.

There are better collaboration tools for 3D design, such as Gravity Sketch, better simulation and modelling through digital twins, and companies like Hyperganic that let designers generate candidates for components automatically based on a set of functional requirements.

The Human Factor

It would be irresponsible to ignore that increased automation will, over the coming decades, pose challenges with respect to jobs. Manufacturing accounts for around a tenth of employment in the US and EU, and likely a majority of these jobs will not survive digitalization.

However, in contrast with, say, the advent of autonomous trucking, the job losses here will be gradual given the incremental adoption, and the basic reality that expensive plant and equipment will typically have decades-long replacement cycles.

This should allow for a more controlled tapering down of some jobs as people retire anyway, and time for others to retrain to work more efficiently alongside supporting robots or, for many, a supportive transition into something else. Like many other sectors, this is a task for society to acknowledge and step up to resolving.

A Bright, “Lights-Out” Future

At Atomico we invest not just for returns, but to support startups that are in some way bettering the world. In this case, higher efficiency, better and more appropriately customized end products, significantly reduced waste and environmental impact, increased safety and better quality of life all come together to motivate us to carry further our early interest in the digitalization of industry.

Furthermore, we see significant advantages in favour of European startups here, given the continent’s strong and proud manufacturing legacy, talent and expertise. There will be challenges, particularly in ensuring incumbent companies work closely with startups, pay for products rather than expecting endless proofs-of-concept, and act as advocates and supporters.

But early signs are very encouraging: Moderating a panel on this topic recently at an event organized by prominent early-stage German VC fund La Famiglia, it was surprising just how much willingness there was on the part of panellists Martin Brudermüller (CTO of BASF) and Matthias Zachert (Chair of the Board at Lanxess) to engage with startup co-panellists from Alchemy and Bigfinite.

We’re excited to help our portfolio companies manufacture this bold future, along with supportive friends like Point Nine Capital, La Famiglia, Entrepreneur First and others. With that in mind, if you’re a startup or an investor that feels the same way, we’d love to hear from you.

spinner