How Engineering and AI will work together.

A friend asked me a question about how his son, with his newly minted MechE degree, would see his job change as AI became more and more relevant to industry. This was my answer:

I was thinking about your question on how engineering and AI will work together in the future, particularly for a new engineer starting out his career. Here is my answer to that, breaking down the five trends that I see in the industry. My answer focuses on industrial applications because most robotics advances are first used by industry and then, much later, by consumers.

(1) Increasing investment cost/wage ratio makes new robotic applications attractive to investors

Since the mid-2000s, the price of computation has been falling, China’s expansion up the manufacturing value chain has been driving down the cost of robots for industrial and retail applications, and the average worker wage around the work has been increasing (particularly in China).

These three trends lead to a large movement towards automating processes that were formerly accomplished by humans.

We see this happening around us:

• User interfaces:
  • Food ordering kiosks and apps that replace minimum-wage till workers.
  • Voice-controlled phone systems that actually work.
• Warehouse logistics:
  • The story of Kiva robotics (now Amazon robotics) is an exemplar.
• Robotics at much smaller scales:
  • Kuka Robotics, a giant German-Chinese industrial automation company, has expanded into small-scale manufacturing with general-purpose robot arms.
  • Miso robotics is a startup targeting the food industry in the U.S., with robots designed to replace low-wage workers. To achieve this, Miso needs to bring the robot depreciation cost down to under US$7/hour, which is incredibly cheap.

Anything that drives the cost of robots down is a promising opportunity. There is already heavy investment in lowering the cost of “hard” cost drivers like stepper motors, aluminum parts, sensors etc. “Soft” cost drivers are yet under-explored.

• Part standardization drives costs down by encouraging competition. (Commercial/hobbyist drones for an example: building a pesticide-spraying drone now costs as little as ten hours of flight time with a spraying helicopter.) As an early entrant into a market, if you can construct a good standard and give it away for free, you can often make a profit off that in the long run. (See Prusa3D)
• Improvements in algorithms/controls can allow greater power or precision from cheaper parts. For example, Trinamic motor drivers allow $20 stepper motors to perform like $100 stepper motors. Advances in antenna and chip design remove the need for a separate antenna – instead, an antenna can be designed to be etched directly onto a PCB.
• Improvements in algorithms allow using cheaper sensors to do more. For example, using commodity cameras for part inspection tends to drive down costs.

(Note: as a general rule, once the expected total cost per hour of using a robot is less than hiring someone, companies tend to move away from hiring to automating. Bigger companies tend to automate first because they can field the capital outlay. Since the marginal cost of automation is much lower than that of a worker, automation is “sticky”: once something is automated it stays automated.)

(2) Advances in machine learning and related fields make new mechanical designs possible

This is one of the ways in which machine learning can influence the discipline of engineering. Thanks to advances in machine learning and optimization, we can create:

• Soft robots. Fragile or difficult-to-handle materials are handled using manipulators that are designed to deform while holding them. This also includes robots that can deform themselves to access inaccessible areas.
• Topology optimization. Given some base geometry, constraints, and a description of the forces on a part, machines can optimize the design of a part to minimize material cost or weight. There are huge opportunities here for dealing with anisotropic materials, especially regarding 3d printing..
• New robot designs. In the 1960s, the first inverse kinematics algorithms made it possible for general-purpose robotic arms to be used in manufacturing cars. With newer control algorithms, we can now control robots with more exotic geometries, like multi-gantry CNC systems.

(3) Advances in machine learning and industrial design make human-robot collaboration possible

One of the biggest barriers to investment in robotics is that robots and humans could not work together, largely due to safety concerns. Unfortunately, it meant that robots were an all-or-nothing proposition, which drove up the cost and risk of automation.

Since then, there have been inventions like safety cages and interrupt sensors that allow humans and robots to work on the same factory floor, but not truly collaborate. Since the mid-2000s, there has been a push to design robots that are safe around humans without separate working spaces and that can truly collaborate with humans. (I was an undergraduate research assistant for Dr. Stefanie Tellex, who specializes in this and related areas.) This involves creating hardware that complies with external forces, control software that limits the amount of force exerted, and general-purpose cameras that grant a robot some situational awareness. (Article about this.)

This is a fledgling field, with many up-and-coming applications:

• Robots in food-service (e.g. Miso robotics) that share fixed production stations with human staff.
• Medical robots that can automate or assist in standard procedures, such as handling the fine details of suturing while a doctor provides oversight, or automatically taking X-rays.
• Consumer robots like automated lawnmowers. These are basically larger Roombas that can hurt you.
• Automated farm equipment, like tractors. Even human-operated tractors are very dangerous; proper safety-aware automation will likely be an improvement over manual operation.
• Automated child- and elderly-care.

One of the key advantages of this is that it permits gradual investment in robotics, which drives the demand for such investment up. Intuitively: small restaurants are more likely to buy just one robot to help out line chefs than they are to replace all their kitchen staff.

(4) Advances in computer science make new production processes possible

Some production processes have only recently become possible. For example:

• Computational hydrographic printing for printing precise designs on curved and uneven surfaces.
• Boeing’s work on remanufacturing aircraft. Boeing has devised techniques to “remanufacture” aircraft. The details of this are secret, but it is suspected that they use engine performance metrics to figure out which parts need to be replaced and which need to be reconditioned.
• New techniques allow robots to use equipment designed for humans. (Miso robotics yet again. I really like their concept!)

(5) Changes in licensing and ownership terms make adopting new technologies cheaper

This is a general industry trend, where it is a lot more common now to lease general-purpose tools and adapt them for your needs than it is to buy or build outright. This applies both to computing (“cloud” services) and manufacturing (leasing contracts with automation companies). This reduces the barrier to investment by reducing risk, reducing the initial capital outlay, and reducing the time to first production. This is likely going to help drive the growth of this field. If I had to guess, manufacturing-as-a-service is going to become much more common and relevant in the future (see PCBs, 3d printing, but on an industrial scale.)

It is an exciting time to be an engineer!