We were promised robots everywhere - fully autonomous robots that will drive our cars from start to finish, clean our dishes, drive our freight, make our food, pipette and do our lab work, draft our legal documents, mow the lawn, balance our books and even clean our homes.
And yet, instead of Terminator or WALL-E or HAL 9000 or R2-D2, all we have is that Facebook is serving us ads that we don't want to click on, Netflix recommending another movie that we probably shouldn't be watching, and Roomba from iRobot.
So what went wrong? Where are all the robots?
This is the question I tried to investigate when setting up my own robotics business (a currently stealthy business called Chef Robotics in the food robotics business) and investing in many robotics businesses / AI via my Prototype Capital venture capital fund. Here is what I learned.
Where are we now?
First and foremost, robots are nothing new. The industrial robot arms with six degrees of freedom (read as six motors attached in series to each other) were actually developed around 1973 and there are hundreds of thousands of them - it's just that so far, almost all of these robots have been the highly controlled environment of factory automation doing the same thing over and over millions of times. And we've formed many multi-billion dollar companies using these factory automation robots, including FANUC, KUKA, ABB and Foxconn (yes, they make their own robots). Go to any auto manufacturing plant and you will see hundreds (or, in the case of Tesla, thousands). They work incredibly well and can take huge payloads - a full car - and have accuracy sometimes up to a millimeter.
More generally, the world of industrial automation is extremely mature and there are hundreds of "system integrators" you can go to and say, "I want an automation machine that performs this use case extremely close millions of times. Build me a system to do it. "This is how Coca-Cola gets its bottle fillers, Black & Decker does its drills, Proctor & Gamble does your shampoo, and more generally how we make most products today. These system integrators can charge you $ 1M and make you wait a year to build the machine, but almost any type of system is possible in this world. The problem with these systems is that they are mostly what is called "hardware automation" insofar as they are mainly mechatronic systems and work disproportionately if the inputs to the system are exactly what they are designed and programmed for; but as soon as you put a two liter bottle of Coca-Cola in a machine bottling designed for half-liter bottles, the system does not know what to do and will fail.
The other big world in which we see a lot of production robots (with the exception of purely software AI agents such as recommendation systems, spammers for email, object recognition systems for your application photo, discussion robots and voice assistants) is that of surgical robots. One of the main players in this space is a company called Intuitive Surgical (market capitalization of $ 66 billion) which has built and already deployed around 5,000 teleoperated robots. Note that these robots are indeed "remotely controlled" by a doctor and are not for the most part autonomous. But since more than 40% of deaths in a hospital are correlated with a mistake made by a doctor, patients pay extra for these robotic surgeries and hospitals buy them en masse; major players like Verb Surgical, Johnson & Johnson, Auris Health and Mako Robotics are following this trend.
What you will notice about both factory automation and surgical robots is that they are found in extremely controlled environments. In the case of factory robots, robots don't really "think", but always do the same thing over and over. And in the case of surgical robots, almost all of perception, reflection and control are performed by a human operator. But as soon as you make the factory automation robots think or the surgical robot makes decisions without human supervision, the systems go down.
So why don't we see more robots today?
The distinction to be made is that we don't see robots today in the everyday world we live in - in uncontrolled environments. Why don't we see robots in the everyday world? What is the main thing that prevents us from reaching our future robotics of the dystopian world? Is this a hardware problem? Software problem? An intelligence problem? An economic problem? A human interaction problem?
To answer this question, it is important to understand what a robot really means. In the literature, a robot is an agent that does four things:
- Meaning: the agent perceives the world using a kind of sensor - say a camera, a LIDAR, a radar, an IMU, a temperature sensor, a photoresistor or a pressure sensor.
- Think: Based on the sensor data, the agent makes a decision. This is where "machine learning" comes in.
- Act: Based on the decision, the agent activates and modifies the physical world that surrounds it.
- Communicate: the agent communicates with others around him. (This was only added to the model recently.)
Over the past 50 years, we have made exponential progress in each of these areas:
- Detection: the prices of cameras and other sensors such as LIDAR, IMU, radar and GPS are falling exponentially.
- Think: cloud computing like Amazon Web Services and Google Cloud Platform has made building software incredibly inexpensive and lets you pay for exactly what you use. GPUs like NVIDIA have been reused from game graphics cards to be able to run parallel processes that are ideal for machine learning applications (and we now have GPUs hosted in the cloud). Algorithms like deep neural networks have relied on the age-old perceptron to be able to do things like recognize objects, understand natural language and even create new content.
- Act: This is probably the most mature area. If we divide the world of robotics at the highest level into manipulation (interacting with the world as we do with our hands) and mobile robots (walking / moving), then the automotive industry has solved most of the problems in the Mobile robot hardware and industrial automation has solved many problems with handling objects (assuming a given pose of the object). We are extremely skilled at making equipment and we have the basic equipment to build robots capable of doing everything.
- Communicate: thanks to the Internet and mobile revolutions of the 2000s and 2010s, we have made enormous progress in the world of user interaction. So much so that today, if we find that a company does not have a simple user interface / user interface, we do not take it instantly seriously. Disappeared companies like Jibo, Anki and Rethink Robotics have made serious contributions in this area.
In other words, purely from a technical point of view (we will come back to economics and human interaction later), it does not seem that detection and action are the main bottlenecks. We have really good and inexpensive sensors and we have excellent actuation technology (mainly thanks to industrial automation).
The problem therefore lies mainly in "reflection". Specifically, according to the dean of engineering at the University of Pennsylvania, Vijay Kumar and founder of the Robotics GRASP Lab, the reason we don't see robots in our daily world is that "the physical world is continuous during the computation, and therefore detection and control, are discrete, and the world is extremely highly dimensional and stochastic. "In other words, it is not because a manipulator can take a cup of tea that he can choose a wine glass. Currently the paradigm for thought that most companies have adopted is based on the idea of machine learning - and more specifically deep learning - where the basic premise is that instead of writing a "program" as in classic computing that takes an input and spits an output based on it, why don't we give an agent a bunch of inputs and results in the form of training data and did it come with the program? Just as we learned in algebra that the equation of a line is y = mx + b, the basic idea is that if we give the machine learning algorithm y and x, it can find m and b ( except on much more complex equations). This approach works well enough for you more path.
But in the incredibly unpredictable world we live in, the idea of providing training data in hordes with the idea of "if you see this, do this" doesn't work; simply said there will be never be enough training data to predict each case over there. We don't know what we don't know and unless we have training data for every single instance that has ever happened to an agent in the past and who already arrive at an agent in the future, this model based on deep learning can not lead us to complete autonomy (How can you predict something that you do not even know possible?). Humans as intelligent beings can really think; agents based on deep learning do not think - they correspond to models and if the agent's current state does not correspond to one of the models which have already been assigned to it, the robot fails (or in the case of autonomous vehicles, plant).
What can we do to make more robots that work?
So maybe the deep neural networks are do not how we get to 100% autonomous systems (which is why companies like OpenAI are investing in reinforcement learning algorithms that mimic a Pavlovian approach based on reward / pain). But while waiting for startups, what happens if the question of how to build a fully autonomous agent is not the right question to ask?
One company that illustrates the idea of not looking for 100% autonomy is Ripcord, a startup based in Hayward, California, which performs autonomous paper scanning. Today, companies have thousands of reams of paper they would like to scan - "no humans have gone to college to become a basic cleaner," says CEO Alex Fielding - and are therefore sending them to Ripcord where the oars are introduced into the cells of the robot which choose and place each sheet, digitize them, then stack them. When talking to Alex in the factory, one of the things that struck me was that he never mentioned the idea of "automating humans". His argument was rather that Ripcord makes a 40x human more efficient. I saw it firsthand - a human supervising four robotic work cells on Alex's premises. In one example, the robot was working extremely quickly through sheets of paper when it perceived a sheet which confused it. At that time, the man supervising the system received a clear notification on a screen with the problem. The human quickly resolved the problem in 10 seconds and the robot came to life for the next few sheets.
So what if the question of how to build a successful robotics business is not "How do you build agents to automate humans?" but rather "How can we build agents to make humans 40 times more efficient while using their intelligence to handle all extreme cases?" As artificial intelligence grows, this seems to be the formula for building successful businesses in the meantime.
Another company that illustrates this is Kiwi Robotics. Based in Berkeley, California, Kiwi manufactures mobile food delivery robots. But when talking to CEO Felipe Chávez, "We are not an AI company; we are a delivery company. "When Felipe founded Kiwi, he didn't invest in a ton of expensive machine learning engineers; rather after building the hardware prototype, he built low-latency software to be able to remotely operate Kiwi. The idea was Originally that humans made 100% of the decisions for Kiwi and slowly they built algorithms to reduce that from 100% to full autonomy. Today Kiwi has a team of dozens of call center operators in Colombia (where Felipe was born) and has made more than 100,000 deliveries. A single human can supervise multiple robots, and the robot makes almost all decisions and humans only correct the trajectory. On the other hand, many competitors who invest independently find it difficult to make even 1,000 deliveries.[[[[ Full Disclosure - I am an investor in Kiwi Robotics through my Prototype Capital fund.]
In these two cases, one of the most important factors is not the machine learning algorithms, but rather the human-machine interface. What is missing from contemporary robotics companies? According to Keenan Wyrobek, the founder of the blood drone delivery company Zipline and one of the early pioneers of robotics, "while the reduced workspace works well for ... business owners in the US market, I have seen countless robotics startups fail with this mindset. . Make sure your design and eng[ineering] the team focuses on the productivity of all users of your system ... I don't care about the quality of your robot, it always has users (people who configure, reconfigure, troubleshoot, maintain, etc.). And if these users are not at the center of your design process, your robots will not function well enough to[n] KING. "
Additionally, according to Amar Hanspal, CEO of Bright Machines and former co-CEO of Autodesk, "The common factor between the two is that robotic companies start with technology first (it's too difficult and somewhat exciting, so it becomes an end goal in itself) rather than the problem they are trying to solve. The key is ... to define a problem you are trying to solve, and then create an excellent UX around it. Robotics is a means to an end, not the end itself. "
What else can we do to see more robots in our everyday world?
So far we've seen that one of the main reasons why everyday robotics has broken its promise is that the world is extremely stochastic and that artificial intelligence based on models based on deep learning is just not good enough to deal with every corner case. So maybe instead of a labor-saving model, robotics companies should adopt the "human augmentation" model. Take the Apple and Airbnb playbook of a human-centered design-first mentality - no engineering - and invest in an incredible user experience.
Here are some other things we can do to bring robots to the fore:
The first is to sell the product before building it. In the software world of Silicon Valley, "The Lean Startup" by Eric Ries popularized the idea of "launch quickly and repeat quickly until you get to the product market." For software startups, this works incredibly well. But with hardware and robotics, what ends up happening is that engineering-intensive startups initially focus not on sales but rather on engineering and they build, build, build. Then they turn to customers to sell, customers say, "It doesn't exactly match our goals," companies don't have enough leads to iterate, and they die. It happened again and again. It seems that for software startups, the lean startup approach works since you can launch most of the time for free (thanks to the cloud), iterate once in the field, deployments are quick and you have five or six shots on target to run money in your seed round. But in the hardware world, you have initial hardware costs, deployments are slow, iteration cycles are slow, and you only have one or two hits.
To be clear, we are extremely adept at hardware; it's just that software-centric Silicon Valley isn't (with the notable exception of Apple and Tesla). One reason may be the lack of sales before construction. Example: Boeing did not approach Juan Trippe, the legendary founder of Pan Am Airlines, and said, "This is a Boeing 747 - do you like it? No. Let me go back and create a new version ... Do you like it now? "(ie, iteration of" The Lean Startup "). Instead, Boeing asked Pan Am to give them an initial order for dozens of units with all the functionality at going ahead so that Boeing can build it right the first time. In other words, Boeing sells its product before it is built. System integrators ask for orders and money before they build anything. the same goes for most computer hardware companies and military branches. Perhaps robotics companies can take a page from the Bill Gates manual and sell MS-DOS to IBM before writing MS-DOS.
One of the benefits of selling before construction is that you can do a consistency check on the economy of the unit. Robotics is one of those areas where there is not only a technical risk but also a unitary economic risk. Many companies have historically found that even if they can come up with a great idea in a constrained environment, develop technology, raise venture capital and build great human-machine collaboration, their economy makes no sense and again they fail. When selling before building, you need to analyze your client's economy as well as yours and make sure it makes sense. If you are trying to sell your product before you build it and no one wants it, it is an extremely safe way to know that your customers are unlikely to buy it, and you may want to move on to the next idea. .
More generally on the economy, we need to move from initial cash flow models to robotics as service models. Many customers who purchase robotic applications have extremely low margins and cannot afford to pay $ 100,000 + in advance for a system (even if the payback is a year or two). Adding fuel to the fire is that the activation energy ends up being too important to change something when they "already have something working." Thus, they reject the product (then the startup dies). We can take a page from the solar / photovoltaic industry here; The economics of solar cells make a lot of sense for many homeowners and yet, for a very long time in the 2000s, we saw very few solar cells. Why? The advance was too much for most Americans, even if the economy makes sense in a few years. The tipping point was not technical but rather financial with companies like Solar City, Sunrun, Sun Power and others innovating on a model where the customer pays almost $ 0 in advance but then has monthly PPA loans where it pays per kilowatt hour that cells generate. The same has been the innovation of cloud computing - rather than buying a group of servers locally to run Oracle and SAP, companies like Salesforce have come up with a "pay for what you use" model. To be successful, robotics companies must do financial engineering so that customers pay very little in advance and pay only for what they use (each hour worked, each sheet of paper scanned, each dish cleaned, each kilometer traveled, each kilo of freight) Shipped).
Another advantage of pre-construction sales is that you can test in the field regularly even if you are also building equipment. Traditionally, this "iteration after deployment" is the advantage of software (compared to Apple, which often begins hardware development for some of their Macs five to seven years before launch). Since you already have a customer, it's in your best interest to make the product work. One of the strategies we have found to be extremely effective is to provide equity capital to your early customers so that they have more incentive to work with you to make the product work economically and technically for them.
But not everything has to be software. Most Silicon Valley VCs now cringe when they see robotics companies that are "heavy in hardware." "We will invest if you take a more software approach," they say, and so today we see robotics companies trying to use almost 100% standard hardware and focusing almost all of their efforts on software. This makes sense in some applications, but the fact is that hardware has failed far less than software and hardware for millennia and we are really good at it compared to the relatively nascent computer age. In many cases, hardware can solve the problem much better than software. Take, for example, garbage collection; today, dozens of startups have raised hundreds of millions of dollars from major VCs to build generic deep learning and reinforcement by learning systems so they can choose and place generic items in a bin. On the other hand, at the PACK Expo in Las Vegas, I got to see a company called Soft Robotics. They adopted a mainly hardware-based approach to picking up trash with a new clamp that, without any computer vision, can pick up and place objects using excellent control (much more consistent than almost all startups based on the computer vision). Of course, building software and training data is important, but why solve the problem more complex when there is a simpler and more robust solution? We shouldn't be running from hardware - we just need to rethink how to do hardware.
More generally, VCs in Silicon Valley have created a mindset that if a business can't be worth a billion dollars, it's not worth investing in or investing in it. The founders of robotics are therefore trying to build a technology that can serve all possible customers in the hope of raising venture capital; and although they alleviate VC, they end up creating a product that doesn't make a customer extremely happy. The best companies to start with had extremely small markets. In our highly dimensional world, trying to build an incredibly generic robotics business on the first day is a mistake. On the contrary, at the beginning, it is important to focus on one (or perhaps two) client (s) in a manic manner. Once you've solved that customer's problem, you'll find that other customers probably want something similar. Robotics probably won't develop as quickly as consumer or enterprise software vendors in the beginning. But it is not unknown. Before Intel and the personal computer era, IT worked very similarly to how automation system integrators work today: you went to an engineering firm for a specific computer that could do one thing - for example calculating the trajectory of your missiles - you pay them $ 1 million, you wait six months and you get your computer the size of a room. Just as computing was slow and not scalable at first, so was robotics. It's okay and there are still billions of dollars to be raised.
Finally, maybe the way to build a successful robotics business is indeed to sell vertical B2B solutions (that is, the "hole in the wall" and not a drill) instead of creating B2C companies aimed at consumers. The latter's promise was simple: if existing customers don't see the technology working for them or the economy makes sense, why don't we both develop the technology and aren't we our own customer? After all, our technology is better so that we can make our own profit and in addition we can control the environment and therefore it should be technically easier too. It was the same argument as the innovative high frequency trading companies who decided to create their own hedge funds instead of selling their technology to other hedge funds. So we saw robotic B2C restaurants, end-to-end law firms that were building AI to automate, and cafes for consumers. The problem was twofold: first, most B2C businesses like restaurants fail and most startups fail, but trying to do both is just too much, especially for a startup with a limited track; and second, many of these brands did not work not because the technology did not work but rather because the consumer brand was not strong enough. The type of team it takes to build a tough technical product is very different from that it takes to build a consumer brand, and often, even if their technology works, the brand was not strong enough and the So customers came once to take a picture, but the retention was not good enough to run the economy. The same goes for education-based and "toy" robotics - although these are "cool", we have yet to see an example of a company that used this model to build a sustainable business, as it seems they are more "nice" to have "than" need to have ". (So when an economic downturn like the one we live in occurs, nobody wants the product anymore.)
There has also been a recent trend towards platforms to allow robotics businesses to facilitate their success, just as AWS has facilitated the success of modern Internet businesses. Again, it sounds great on the surface, but the difference is that before AWS, there was a thriving set of software companies that built big companies and had the money to pay AWS for a better product. But today, there simply aren't enough robotics companies that have enough revenue to make sense of these B2B companies. It always seems like we need the iPhone "killer app" before the App Store platform makes sense.
Areas prone to disturbances
In other words, we have a long way to go to see the robots in our everyday world, because there are so many places where robotics companies can go wrong. Here are the types of robots that I think we will see more of in the world on a daily basis in the short term (the next two to four years):
More autonomous factory automation. For industrial automation, customers already exist. If we can build better technology that makes these systems more self-sufficient, we will see many more customers who want it.
Semi-autonomous and remote-operated companies. Semblable aux robots chirurgicaux, Tesla pilote automatique et Kiwi, nous verrons beaucoup plus d'entreprises dont l'objectif est l'autonomie partielle et d'augmenter les humains sans les remplacer.
Robots basés sur la manipulation dans des paramètres de type usine. En 2015, principalement en raison de l'investissement de Google dans les voitures autonomes, les VC ont investi des centaines de millions dans des véhicules autonomes en partant du principe que "conduire c'est conduire c'est conduire". Si nous pouvons résoudre la conduite pour une voiture et dans une ville, cela peut probablement évoluer assez bien. Aujourd'hui, nous sommes dans un peu d'hiver dans les véhicules autonomes et très peu d'entreprises semblent avoir une idée de ce qu'il faut faire ensuite (principalement parce que le monde est tellement aléatoire et que l'apprentissage en profondeur peut ne pas être suffisant). D'un autre côté, la manipulation a été laissée de côté et semble aujourd'hui faire son retour alors que nous voyons des ingénieurs quitter des constructeurs de véhicules autonomes et chercher quelque chose de nouveau qui pourrait en fait être en production plus tôt. Les applications de manipulation ont tendance à se trouver dans des environnements extrêmement contrôlés et nous en verrons probablement plus (comme les microfactoires de Bright Machines et les robots de tri de recyclage d'AMP Robotics)
Dans la même veine, il existe aujourd'hui une tendance à "évoluer vers le cloud". Imaginez qu'avant la première révolution industrielle, nous fabriquions des textiles dans nos maisons. Mais nous avons alors réalisé que nous pouvions centraliser la production de textiles dans les usines et profiter des économies d'échelle. En conséquence, nous voyons aujourd'hui très peu de personnes fabriquant des textiles chez nous. En appliquant cela à aujourd'hui, si vous imaginez un monde dans lequel presque tout se déplace vers le "nuage" et que vous envoyez vos tâches ménagères à quelqu'un d'autre pour les faire en utilisant une installation robotique centrale (cuisine, vaisselle, lavage de linge, pliage de tissu, etc. ), il existe une énorme opportunité d'appliquer des robots qui affectent la personne ordinaire mais qui se trouvent dans un environnement où les robots fonctionnent le mieux (usines).
Peut-être que la seule chose que nous ferons dans nos maisons sera le nettoyage, et donc il y a et il y aura toujours une énorme opportunité pour nettoyer les robots des systèmes pour nettoyer les maisons intérieures, tondre les lois extérieures, nettoyer les centres commerciaux intérieurs et autres applications B2B, et labourer neige extérieure.
La robotique est toujours très prometteuse et certainement réalisable. Vendre avant de construire, s'assurer que l'économie de l'unité fonctionne tôt avec des paris à faible risque, tester le système souvent sur le terrain, fournir aux premiers clients l'équité du conseiller pour aligner les incitations, construire un produit pour résoudre un problème pour un client particulier plutôt que de construire quelque chose de générique , pensant aux robots comme une combinaison de matériel performant and d'excellents logiciels plutôt que des logiciels seuls et la poursuite d'applications B2B verticales peuvent aider. Mais dans un sens plus large, plutôt que de frapper chaque ongle avec le même marteau de mentalité logicielle, il est peut-être temps de penser à partir de zéro.
