Within a week of unveiling the much touted Model 3, Tesla Inc. received reservations of over 325,000 cars at an estimated worth of US$14 Billion. To ramp up production at its Fremont factory from Zero to the proclaimed 20,000 cars/week target that Tesla said it could achieve by December 2017 would be a nightmare even for the most seasoned production engineering teams, and most auto giants expressed disbelief. For the upstart however, the potential of automation was a glimmering ray of hope. Tesla embarked on a blitzkrieg of investments in automation.
“Robots will be able to do everything better than us”
-Elon Musk, July 2017
…proclaimed Elon Musk in July 2017 when the first Model 3 rolled out of the assembly lines.
Less than a year after the statement and the first roll outs, Elon was forced to retract his position on automation, admitting that he underestimated humans and that excessive automation at Tesla was a mistake. Production barely touched 2000 cars a week, Tesla had to revert several processes back to manual to ramp up the production. Eventually, the plant could hit its revised target of over 5000 cars/week by the second half of 2018, still a far cry from what was originally hoped for, and not without some handy helping hand from the good old factory worker.
“Yes, excessive automation at Tesla was a mistake. To be precise, my mistake. Humans are underrated,”
- Elon Musk, April 2018
In case one thought that Elon’s less than desirable tryst with robots had something to do with the nascency and naivety of a brash new player in manufacturing, even seasoned manufacturers seem to have had their share of disillusionment. Adidas famously aspired to manufacture shoes in hyper-automated factories in the US and Europe to reduce reliance on Chinese factories, but only 4 years after, has had to shut down these factories in favor of shifting back to Asian factories.
Elon Musk, Adidas and anyone else who has experienced first-hand an attempted transition from manual labour to full automation, would likely envy the relative simplicity with which humans can be trained to perform a plethora of manual labour tasks in comparison to the nightmarish complexity of dealing with today’s automation technology.
Manual labour is simply about picking, orienting & placing randomly occurring objects
When observed closely, every manual labour task in any industry is simply about manipulating objects in the real-world – one learns the basic physics about objects – such as their colour, texture, size, shape and mass distribution, observe how they occur in 3D space, and develop simple strategies to pick, orient and place these objects as required. Yet achieving the same in automation even in highly structured environments like a factory floor, involves months of planning, systems integration and prolonged testing and stabilization of hundreds of sub-components even to achieve a simple task.
Traditionally, automation looked myopically at specific tasks and implemented strategies to perform these tasks autonomously. Moving an object in a linear manner from a known location A to a known location B would involve a set of motors, sensors and guiding mechanisms. This strategy however quickly escalates in complexity as we add more complex goals while manipulating objects, or when object geometries become more sophisticated. As we seek to add variability in the initial and final location and orientation of the objects and the intermediate path to take, the complexity of designing an automated manipulation system exponentially increases.
As the location or orientation of objects at either ends of the task to be performed becomes variable or unpredictable, we enter the realm of tasks that are impossible to automate today.
Predictability is a pre-requisite for automation
For predictable tasks, originally systems were implemented as special purpose machinery, integrating several motors or pneumatic actuators, mechanical components, sensors and controllers. The cost, complexity and lead times in designing, manufacturing and validating each system would be enormous, even for basic tasks.
As automation technologies evolved, the need for a reconfigurable object manipulation system emerged. This led to the invention of articulate multi-degree of freedom object manipulators, more colloquially referred to as “industrial robots”, which have been marketed since the 1950s and saw a major boom in the 1970s with the likes of ABB, KUKA and FANUC driving the adoption.
A $48B Industrial Robotics Industry Ripe for Transformation
Since then, Industrial Robots have steadily grown to a mammoth $16B industry, with an allied ecosystem industry of services and components worth at least twice that amount, as per International Federation for Robotics.
Today’s industrial robots have 3 to 7 internal axes of freedom and ability to work in coordination with several more external axes using a centralized controller. With position repeatability as low as 0.01mm, speeds reaching 3m/s, and payload capacities of several hundred kilograms, one can safely conclude that these beasts dwarf the human arm’s ability by miles.
Yet, the true robot revolution is still ahead of us.
The robot automation market size still pales in comparison to how much the world spends today in employing manual labour. A McKinsey research estimates that US alone spends about $1.3T in wages associated with manual labour.
On closer observation, it is easy to recognize the gap. While robots as mechanical manipulators are great at reproducing any required motion profile faithfully, they still lack the ability to adapt dynamically to changing conditions. This means, one can deploy robots today only when initial location and orientation, final location and orientation, and the path between the two locations is always known and predictable.
The large plethora of tasks required to be performed in industry by nature involves objects whose location or orientation in space may not be inherently predictable. Without cognition, even a few millimeters or degrees of variation in the location or orientation of the objects renders a robot incapacitated for reliable performance. This adds tremendous pressure on automation system designers to build customized structuring mechanisms in their automation processes with sub-millimeter tolerances just so that today’s robots may “blindly” pick and handle objects.
The challenge turns from difficult to impossible when handling complex geometry objects, which might not have stable bases to easily structure using vibratory feeding mechanisms, or those that might even get entangled in a bin condition. A connecting rod is one such component that is extremely difficult to organize from a bin and in most automated manufacturing lines, this initial structuring is still done manually. The seemingly simple task for a human involves a level of sophisticated cognition yet achieved in automation.
Picking a complex geometry part, orienting and placing as required involves a level of cognition of the objects and the scene that Industrial Robots of the day don't possess
As we move up the value chain, in assembly where a variety of objects must be handled, it becomes infeasible to custom build structuring mechanisms for individual components at each stage. Wires and cables for instance are impossible for robots to handle as the very geometry of the object itself varies from time to time.
Picking a screw and accurately placing in a tapped hole and making the first few turns such that the threads don’t slip, and jam is a task robots haven’t mastered yet, and just imagine how pervasive the simple screw is. Practically every object we assemble and manufacture today have screws in them.
Robot deployment statistics point that robots till date have seen very limited penetration in terms of applications – most robot installations are in fact for painting and welding jobs that do not involve any sophisticated object handling at all.
A walk through at any advanced car manufacturing plant would reveal the same. Elon Musk talks about this challenge in a walk-through video of his plant.
Despite the purported universality of industrial robots, a typical robot system still requires painful system integration of several subsystems and components to ensure repeatable performance. The industrial automation ecosystem is still vastly diverse, with several hundreds of companies playing their piece in the larger puzzle. A universal automation system that can work for all kinds of objects and tasks while handling unpredictability of location and orientation of objects and dynamic alterations in required motion paths is still elusive.
The intelligence of objects must come first
At the heart of this challenge ahead of us is a deep-rooted, often unacknowledged, perspective in solving for industrial automation, that is a baggage of legacy we carry with us.
Automation has always and continues to look at the problem from the task’s perspective – how best do we achieve the task of moving an object from location A to location B with precision and repeatability. The focus has until now simply been on bringing universality in implementation of manipulation tasks and not on the universality of understanding objects from the perspective of what all ways one could manipulate them.
Why, every automation system deployed on the planet today is simply “unaware” of the object it handles. While this simplistic strategy is enough to ensure repeatability in cases where they are painstakingly implemented in a fully controlled rigid environment, it cripples the ability of the system to adapt to even the slightest of variability that inherently occur.
In a world most assuredly advancing towards hyper-consumerism of commodities, that too with the expectation of instant-delivery of goods and rapid evolution and revisions of products, can automation scale with demand, relying on such primitive rigid legacy strategies?
Automation methodology has aged, and we now need a paradigm shift.
