Autonomy startups are such a mess

AI startups just use data logging and robotics startups are all in dark mode: where are we heading? I just want to hear what people are really doing, and then I can decide its value.

The dream so many people have: take a cutting edge idea, turn it into a product, and improve tons of lives (with a cash return along the way). I’ve said this, I know people who are starting companies, and I know plenty of people trying to join similar companies. Whenever I get deep into the weeds of the area, it really seems like companies that work exclusively around the areas of artificial intelligence, robotics, and autonomy are really messy. In this post I talk about some reasons this could be so:

  1. AI startups don’t do AI. I think this is driven by venture capital and public misunderstanding.

  2. Robotics startups are really weird and hidden (I give this point of view from the lens of restaurant robotics).

I conclude with my thoughts on making a robotics startup, or joining one right now.

We’re burning up out west.

AI Startups & Not Doing AI

How did AI-startup = data science hype machine become the baseline when I hear about a new AI company? It’s pretty hilarious. That being said, riding the wave of deep learning with a unique dataset can be extremely profitable.

These are some of the hottest names in AI (or should I just say data science) when searching for this article. All of the .ai domain names represent a non-negligible portion of Anguilla’s GDP [the island].

Data is what makes the largest technology companies so valuable. A good estimation is that Google knows your preferences better than your parents. Facebook makes over $26 a year off it’s most active users. This is all from having the best data on the web — they know your location, your spending habits, and your concerns. Data access and control are driving the technology landscape.

Zooming in

Let’s look at some of the most valued “AI” Startups that raised more than $100million in 2019. I did some digging on each, and I’ll reframe their application as a new data access solution. Even what the companies self-advertised as is intriguing.

1) VacasaVacation Rental Management

This one doesn’t even sound like AI at all. This is the Airbnb for creating new home rentals. They have all the data to know what properties to build. They claim “AI-driven tools make it possible to deliver this consistent experience in unique vacation homes around the world,” but realistically I would guess this is basic regression and decision trees.

2) SamsaraIndustrial IoT

They should say — web-connected sensors. This is a great business, they are going to sell a ton of products to people looking to broaden their data pipeline. Smart cameras give rich data and can connect to the cloud. It’s all in the first sentence of the story, except for how AI will help them provide products, “We believe that inexpensive wireless data, AI, and an explosion in camera technology has made it easier for organizations of all types to gather data in ways not possible before.”

3) TripActionsCorporate Travel & Expense Management

I included this one to prove that some of the AI company press frenzy is just chaos. There’s no mention of AI on the ‘about’ page or automation for that matter, just “A travel platform that takes care of business and your travelers.”

4) ThoughtSpotSearch & AI-Driven Analytics

This company takes enterprise data (emails, logistics, etc) and finds trends for them. The AI-Driven insights will provide a service so good that “anyone can find insights hidden in their company data in seconds.” Another example of give us your data, we will do some analysis on it, and make you better. I am sure it works for companies that don’t have time to do the analysis themselves. The best part is middle-man data companies get better with every extra costumer.

5) CloudMindsOperating Smart Robots for People

This is a startup that wants to bring affordable robots to homes by 2025. I’m onboard — great to find it early. The motto: “we can make Robots as smart as a human by using a cloud brain.” Sharing data isn’t necessarily a cutting edge innovation in AI, but you get the point. Again data comes up, but in this case, it’s data sharing.

6) IcertisEnterprise Contract Management in the Cloud:

Icertis is the leading provider of contract lifecycle management in the cloud. Not to beat a dead horse, but I think the value gained here is mostly from having the most data.

7) SparkCognitionCognitive AI

This is the first one to come up that seems like it might be doing some impactful AI research — “We build artificial intelligence systems to advance the most important interests of our society.” Consulting-based AI companies are going to be big. I’m curious if any of the big names in consulting swallow them up before they start taking market share.

8) Vectra AIAI-driven Threat Detection and Response Platform

Haven’t heard of this one, but Vectra is “the world leader in applying AI to detect and respond to cyberattacks,” according to themselves. I’m guessing this is some cool time-series regression and a lot of data. Could be some awesome machine learning here, I will give them credit. Right on.

9) ScaleThe Data Platform for AI

Data labeling services ultimately. I have met multiple people at Scale, they all seem great, and I give them my seal of approval. They took what used to be the bane of graduate students, data management, and made it seamless. This company is going to be around for decades (or get bought soon).

10) AutoXDemocratizing Autonomy (haha)

When you encounter an autonomous vehicle startup, you should always ask yourself, “how is this one different?” The claim here is “building the safest AI driver,” but I am not sold from their website. Time will tell. The only autonomous vehicle company doing impactful AI now is Tesla — they have end-to-end learning trained on billions of miles. I’m sure a couple will catch up

Linking the graph

I am a little surprised how little of them are AI-forward considering the press they get all involve computer intelligence. AI will bring in the eyes, and when you deliver a product on useful data, no one (especially investors) will think twice. Most of the companies have a clear data-leverage, and a couple of them have a problem space that seems truly innovative.

I sourced the most recently funded startups from this article. I had to remove a few of the companies because they’re not intelligence startups, but they’re digital and making waves. The source for most of this information on startups: Top 25 AI Startups Who Raised The Most Money In 2019.

Source — the author’s old day job.

True AI Startups are Coming

We can see the wealth created by these companies just by having access to the data. In the next decade, innovations in machine learning will be regular and impactful. The 2010s was the emergence of big data and it’s boons — I think the 2020s will be determined by who has the best AI. The best AI will get the most consistent views, it’ll have the lowest logistics costs, and much more.

Machine learning breakthroughs and data infrastructure in the last decade are setting the stage for the 2020s to be filled with truly AI-leveraged unicorns.

I am worried that the looming recession will cause increased aggregation in the technology industry. The economics are simple, but the more these companies know, the more they’ll rewire our brains to make money.


Can robot chefs re-open restaurants?

Looking at robotics startups through the lens of a relevant industry: restaurant automation.

My big question: did these companies enter the market thinking they will be niche robot sellers, or how did they end up here?

Is this a VC problem, an incentive problem, or a technology problem? The companies I outline below all seem way too specific. Why make an individual robot you try to sell to individual restaurants. Autonomy is a problem of scale and restaurants are an art. Art’s aren’t always meant to be scaled (and you cannot force scale upon the best ones), and you cannot finance true autonomy startups on niche applications alone.

A survey on the robotic restaurant industry.

Robots? Restaurants? This would’ve all seemed too sci-fi in 2019, but we are about to skip a decade of automation adoption due to the coronavirus. The future is here, and we no longer need humans to get us fries in a drive-through — so these companies claim.

What are the technologies that enable robotics kitchens? These companies all look like framing classic problems in robotics in specific contexts related to the foodservice industry. They don’t need to push the limits of the research frontier — they need to put research in the right place.

What does this mean for the food we get? I think restaurants that adopt these will take a 10% hit on peak flavor and experience for the potential of 100% consistency (the restaurant equivalent of a Jura).

Pick and place — a reasonable challenge

Pick and place is a classic problem in robotics. I’ve heard multiple of my colleagues rant at the boredom of this problem, but it comes up everywhere: logistics, manufacturing, delivery, cooking, etc.

You have an object in position A, and it needs to move to position B.

This ultimately is what 90% of cooking at scale is. Ingredient i needs to go into dish j at time t. I don’t think there’s a good mechanism for robots to feedback based on taste (e.g. the magic of salting a dish from 0 to perfect and getting many flavors to emerge), but I do think robots can repeat these tasks.

I for one would be totally happy if a robot was making my sandwich at a subway, maybe I am just a clean freak. (Not that I eat at subway, but you get the point.)

These two companies are so similar. They both are robot packages that the end user reprograms to accomplish multiple tasks (I’m pro Dexai because I saw their robot with an ice cream scooper first). I’m just surprised they both end up here, what’s to stop a bigger organization from investing 10x in a slightly more broad robot, that can come and swoop into their market? Is this possible?

Interested in pick and place? Read more about it here, or maybe watch this video, or check out my colleagues at Berkeley. What makes it hard:

  1. the infinite number of potential objects.

  2. the infinite orientations of said objects.

  3. humans are really good competitors.

Planning and routing — a safe option

This is a solved problem. I remember watching talks in my internship at Facebook about how the robots can navigate spaces without a map: they just need a goal. When you give a robot a map, it’s an easy problem. Aside — This seems like an amazing way to keep social distancing rules intact.

Strap the robot with redundant sensors, tell it to stop moving when an unknown object is too close (customer), and get to a table to deliver food. Remember: a restaurant owner can scan their floor-map and upload it to the robot, so it knows exactly where to expect chairs, tables, other robots, etc.

Does this mean my food will never get left out to chill off on the kitchen-server border? Sadly, no.

This company makes tall Roombas / moving barstools to deliver your food. It seems like an easy solution, but my question is why do they do it better than a logistics company that has been deploying the same robots for years, at scale?

Dextrous manipulation — far out

Scroll time a tab on this website and you see a complex robot hand. Either this startup has some research it isn’t sharing, or it’s trying to build hype. Robots are just starting to solve tasks like Rubik’s cubes and finger twirling, and they expect it to soft beat eggs with a whisk?

I do think there’s a time when this will all be doable, but having robots do all the dextrous tasks of cooking is not on the menu for the next few years (I think it needs a decade).

There isn’t much information on this one, but it seems to be a dextrous manipulation robot for cooking. I doubt they really are making advancements on dextrous manipulation, but if they are, why are they just doing cooking? That is one of the big open problems in robotics — matching the wonder of the opposable thumb.

Some other reporting and related areas.

There have been many attempts to capture the novelty of robotics by opening human-less restaurants or coffee shops. The automating of the food to the table takes more coordination than a few robots in a restaurant. Looking at this in the context of the supply chain could be the more valuable venter — figure out how to automate and optimize the supply chain with humans out of the loop.

I used to walk past a robotic coffee shop on my commute from Facebook SF to my brother’s office: I never once saw it working.

I love the takeaway of the last article that “maybe we still want to be served by humans,” which I think is a neglected point as we push to automate. I think we need to make sure automation is doing good for all (physical, emotionally, and more) before we just pour money into the checking accounts of a few technology brands.

Photo free from Pexels, with some visitors from well-funded robotics startups.

My Take on Robotics Startups

I think I would be happiest working at the right startup: the intersection of a tangible problem to solve, interesting day-to-day work, and fantastic people. Finding that intersection is so impossible that I will probably avoid it for the next few years. Hopefully, I can either just make my own that solves these problems, or the field will change.

The stealth-curtain cult

Robotics startups are a little culty in how they like to advertise, recruit, and exist. Stealth mode means they don’t share what they are working on publicly, and I think this hand is forced upon them because when someone sees a robot, they may just be able to make a new one themselves (the anti-competitive advantage of robotics companies). Here are some of the problems I have encountered:

  • My friends who interview for robotics startups, even into the late states, never really know what the company is doing. We get it, you want your hype, but, eh.

  • Hearing about trying to start a robotics startup is janky: you want a proof of concept early for a task that entails 90% embedded software and then you need to scale the product to fleet-scale. In between this is when a big company comes knocking to acquire you to round out its portfolio of niche companies.

  • Robots are too far away from being personal computers (general-purpose devices), so there is no pitch that is both not bullshit and truly unlimited in scope. I think the startups realize this and think it’s easier to say “we will do this one task the best,” without realizing it could be long term limiting.

Platforms or tools

Too many robotics startups these days make physical tools. If you make a tool, you’re competing in a small subset of tasks. If you make a platform, you’re competing in every task where you can think of a tool for doing it.

The big robotics companies of the future will be making platforms. Waymo is taking this approach while they learn about self-driving: they are making an autonomy stack that’ll work in many other applications. They future Waymo product will be: choose your sensors and budget, and we offer you ABC performance metrics. That model is so much more compelling than, you should buy our robot set at this price tier for this specialty. Let the customer bring you the agent (robot/device) and you provide it with the learning tools to succeed.

I like the approach of Covariant: computer vision tools for autonomy tasks, with analysis of performance et. al setting up for reinforcement learning (I still have always heard they are not deploying reinforcement learning, yet). They have warehouses come to them, choose from a set of robots from external providers, and modify it with their AI platform. The aggregation model of robotics.

What would I make

With my expertise in model-based reinforcement learning and control theory, I would probably target on-line controller improvement for applications that have high failure rates right now. Reinforcement learning isn’t replacing classical controllers anytime soon because it lacks performance guarantees (also discussed as robustness), but if it can create a controller for an autonomous agent that a human used to have to do, that’s big bucks.

My technical expertise would be in understanding state-action spaces, dynamics constraints (high-level), and so on. Getting hired because you know how to work with one specific robot or one specific tool is just limiting your growth (and you have to compete with everyone else that claims to be an expert on that tool, and competing is hard). I’m pro uniqueness, and against direct competition.

More recent writing from the author:

Author bio. Tweet @natolambert, reply, or subscribe here. Thanks for engaging.

Leave a comment