Robots and automation will take over jobs. The world’s heading towards an era of jobless growth. It’s not just blue collar workers but lawyers, bankers and even journalists are on the firing line. Doomsday prophesies about artificial intelligence are many.

A Palo-Alto based startup, with its engineering office based in Bengaluru, recently raised $10 million in funding betting against that premise. Meet Drishti, a computer vision startup which wants to be the ‘Google for actions.’ It uses computer vision to recognise and record human actions on the factory floor.

“What we’re solving is a 100-year-old problem, it’s a 12 trillion dollar problem that all of manufacturing deals with,” says Prasad Akella, the CEO and co-founder of Drishti.

The problem statement? How do you build the best production system so that it delivers productivity, quality, and things like traceability, which are very fundamental parts of the manufacturing process?

Drishti was founded by veterans from the tech industry in 2016: Akella was part of the team at General Motors that created the world’s first collaborative robots (cobots), now a multi-billion dollar market. The chief technology officer, Krishnendu Chaudhury is a computer vision and machine learning research scholar, previously a principal scientist at Flipkart and has had decade-long stints at Google and Adobe. The third co-founder, Ashish Gupta, was the co-founder of Junglee (acquired by Amazon), and Helion Ventures, a venture fund based out of India.

“We hope to be the Google of action, going forward. In other words, what Google did to generic information, we are doing it for digitised data sets of human actions,” says Chaudhury. The whole case for the startup is built on the assumption that humans are going to be around.

FactorDaily spoke to the founders about how they stumbled upon this problem statement, what Drishti does to solve the problem and counter-intuitive bet on why that humans aren’t going to be replaced in the factory floor. Here’s a closer look.

Drishti’s homepage has a slick video presentation which outlines their mission statement – it revolves around solving a 100-year-old manufacturing problem – that’s around the time when stopwatches were introduced to measure assembly-line human labour productivity and efficiency at the factory floor.

While this brought about a leap in efficiency, the approach is old in a post-IoT world. About 90% of the factory work is still performed by humans and their activities are not available for any data-driven analysis. Drishti’s central thesis is that converting human activities to data at scale could boost productivity, factory margins, and employment in the age of AI.

Using data from the factory floor to improve factory output has been attempted before, most famously by industrial giant General Electric. The company’s Predix IOT platform uses data generated from machines. But that didn’t take off as planned (more on that here). Unlike GE, Drishti’s solution doesn’t rely on sensors as much as it does on computer vision.

Whose idea was it though? “I have the dubious distinction of having done that part of the work,” says Akella. He worked on the idea for Drishti during his year at the famed SRI International as an entrepreneur-in-residence. Many seminal inventions that shape the world of computing today — from the mouse to LCD displays, and optical video disks — came out of the Institute. Drishti raised a seed round on the back of the prototype.

Akella referred to a recent column in which he poses three interrelated hypotheses that highlight how humans aren’t likely to become obsolete anytime soon. In his column, he argues that there are currently more than 340 million temporary and full-time manufacturing workers as per a report by Goldman Sachs Research and not all of them are going to be replaced by robots. The global robot population is predicted to rise from 1.8 million in 2016 to 3.0 million by 2020 and it will only displace 6.7 million jobs, a small fraction of the overall manufacturing workforce globally.

What Drishti aims to do is take the technology that’s making robots attractive to manufacturers – namely machine learning, computer vision, big data – and apply it to augment humans. Their technology gives manufacturers data-driven analytics that can help improve workforce productivity, reduce costs on quality checks, and improve traceability by making human actions on the factory floor searchable.

Chaudhury shares a cricket analogy to explain Drishti’s product offering – “What if we could analyse your batting style and tell you how are you functioning against a pace bowler or spinner? You can do that automatically, and point you to the right parts.” AI researchers at IIIT Hyderabad have, in fact, done research work in the field of sports analytics, exploring the potential of computer vision across a variety of sports – from tennis to badminton.

“If you’re a worker on the line assembling a part of a TV, if I could point to you exactly better ways of doing it, then that becomes very useful to improve your own effectiveness,” says Chaudhury.

Warehouse workers at Amazon are famously known for being tracked all the time and peeing in bottles. Would factory workers warm up to a camera pointed at them all the time, and could it make their work conditions more oppressive? Akella begs to differ, explaining the process by which it is introduced into the factory floor. “We actually ensure that a couple of the associates from the assembly line are a part of the team. The reason we’re doing that is is twofold – one is because they are the domain experts. The second is that they recognise immediately that our focus is not on the person, it’s on the process,” he says.

“The minute they realize, this is no different than not another poka-yoke system, (a Japanese term that means “mistake-proofing”). It’s another technique to keep track of the process on the line. It’s the next generation of that kind of technology, and it all ties back to the Toyota production system, which gives authority to the workers on the line. It’s all about empowering the workers,” he adds. “And so for us, empowering the worker, extending the human’s capabilities are central to what we do.”

Building an A-Team

Gupta, who knew Akella from their graduation days at Stanford, says that he saw a once-in-lifetime opportunity to work at on a problem that was attempting to transform an industry that’s measured in the trillions of dollars and reshape the relationship between humans and technology.

“I jumped at the chance to work with him when he started Drishti. Prasad is unusual in the diversity of experiences that he brings: software, on the one hand, manufacturing on the other, and management and deep science on the third. This combination makes him uniquely suited to this company,” says Gupta.

“Chaudhury was introduced to me through the Bengaluru Google community. I first met him to invite him to a panel on AI at the TIE Global Conference,” he adds.

The prototype developed in 2016 needed a lot of work, as it wasn’t a cleanly functioning prototype. Choudhary helped build out a deep learning prototype by late 2017.

“One of the problems you have with startups is a high failure rate. So if you think about it, you require multiple sorts of levels of domain expertise to play,” says Akella. Drishti is solving the problem statement in three dimensions – domain, the technology, and building a business, and each of the co-founders lends themselves to them – Akela brings in domain expertise in manufacturing, Chaudhury brings decades experience in deep learning and computer vision, and Gupta brings in business expertise.

“What we’re trying to do is try and improve the odds, taking on a big problem with a fantastic technology, with the knowledge of how to build companies,” says Akella.

The Drishti team is presently a team of fifteen people, with a bulk of the developer team based out of Bengaluru. Besides developers, the startup also employs a large number of labellers to generate training data for Drishti’s machine learning systems. Their careers page lists openings in Bengaluru for deep learning, product, and software engineering roles.

“There’s a lot of manual work in the industrial automation space, there is umpteen number of workflows that can be automated using AI or even without using AI,” says Manish Singhal of AI focused early stage fund Pi Ventures. “From a uniqueness perspective, there are so many use-cases in industrial IoT, that you will find a lot of companies doing different things – it’s a very wide panorama. I think the nature of jobs will change – in any automation or industrial IoT use-case, it will impact somebody’s job description,” says Singhal.

Current customers include global automotive and electronics manufacturers, though the founders declined from naming them. A bulk of their research and development work is being driven out of their engineering office in Bengaluru, says Chaudhury, adding that they have filed patents for their research work around computer vision and action recognition.

---