Google Nest Learning Thermostat 4th generation

Google's Nest Learning Thermostat is still the one to beat

Google Nest Learning Thermostat 4th generation

Image Credits: Brian Heater

A refurbished third-generation Nest Learning Thermostat was one of the first things I picked up upon moving into my first house last month. As a longtime New York City apartment dweller, I’d largely observed the device from afar, wondering how a thermostat of all things became foundational to Google’s hardware play.

My timing could have been better, of course, with the fourth-gen device just now launching. But as a financial adviser told me before I settled on this place, “the best time to buy a house is last year.” If you sit and wait for the perfect moment, you’ll be sitting and waiting forever.

During this especially brutal summer, I’ve come to appreciate a lot about the Nest — though buying one used on eBay wasn’t the smartest, as the thing could never stay connected to Wi-Fi. Aside from access to central air, the greatest thing that the installation afforded me is confidence.

Live in an apartment long enough, and you’ll begin to doubt yourself when it comes to even the simplest tasks. On one particularly sweltering mid-July afternoon, however, I flipped off the house’s power, removed the ancient digital thermostat and wired up the Nest. The only thing more surprising than me not putting myself in the hospital was the fact that it worked (Wi-Fi glitches aside).

After the thermostat, I undertook a few other small wiring tasks and — as you can probably guess reading this — did not kill myself. When the new Nest arrived at my front door, I unboxed it and had it up and running in a few minutes. Again, still alive. Your mileage will vary, of course, depending on several factors, including your HVAC system’s compatibility.

For me, it was as simple as cutting off power, popping off the old face, disconnecting the wires, swapping out wall plates and popping the new model in. The Google Home app walks you through the relevant steps. After connecting to your house Wi-Fi (it worked this time, thankfully), it’s ready to go.

Image Credits: Brian Heater

The fourth-gen model does have one extra step, as it ships with a second sensor. The pebble-shaped accessory can be placed on a flat surface or attached to the wall anywhere in the house away from the thermostat. The more you add, the better idea the Nest has of overall temperature. It’s a nice in-box addition, and you can always add more as needed. My place is small, so I should be good with the one.

The new thermostat also has a picture of what the weather is like outside. This, coupled with the device’s onboard sensor and any other secondary sensors, informs the system’s goal of achieving the user’s preferences without wasting electricity. Google gives the following examples: “If it’s a sunny winter day and your home gets warmer on its own, it will pause heating. Or, on a humid day, the indoor temperature may feel warmer than intended, so the thermostat will adjust accordingly.”

That’s essentially the heart of the Nest Learning Thermostat. The system utilizes AI to analyze the data it collects. From there, it follows your patterns and adjusts to fall within your temperature comfort spectrum — hence the “Learning” part of the name. It’s impossible to give the Nest a full review at this early stage, as proof of efficacy is only borne out over time.

Beyond ease of installation and use, one thing I absolutely can say right out of the box is that — much like its predecessors — it’s a great-looking piece of hardware. I suspect that prior to the first Nest Thermostat’s arrival in 2011, few homes gave much thought to how their thermostats looked. It’s not a coincidence that two of Nest’s founders came from Apple. The startup effectively made the iPod of thermostats, right down to the scrolling navigation.

Image Credits: Brian Heater

The new model’s 2.7-inch display is 60% larger than its predecessor’s, now stretching edge to edge. In fact, the screen is larger than the base, creating a kind of mushroom shape when viewed from the side. Google gets more use out of the new curved display, utilizing a proximity sensor to note when someone is in the area and displaying a range of different faces, including different clocks, the weather or just your standard thermostat info.

The thermostat is the first thing one sees upon entering my new house. It’s right there on the opposite wall and, for better or worse, it immediately draws the eye. It’s a good spot for a display that showcases additional information. It’s not a smart display by any means, but the additional info is a nice touch for anyone passing by who isn’t obsessed with the indoor temperature.

The Nest Learning Thermostat remains a beautiful piece of industrial design, with some fresh twists that help it more seamlessly blend in with other smart home devices and the home in general. Speaking of twists, the company has maintained the rotating interface, with the tactile clicks that drew many to the line in the first place.

Image Credits: Brian Heater

Google is hoping to reinvigorate its smart home offerings for the age of generative AI-powered assistants, so it might as well have returned to the device that started it all. The fourth-generation thermostat perhaps isn’t exactly the sort of top-down reinvention one might have hoped for after nine years between devices. But then, despite an influx of competition in the smart thermostat segment, the previous model continued to hold its own.

It’s a $280 device, which is admittedly a big investment for a product few thought much about a decade ago. But it won’t be coming off my wall anytime soon.

Boston Dynamics... robot down!

Humanoid robots are learning to fall well

Boston Dynamics... robot down!

Image Credits: Boston Dynamics

The savvy marketers at Boston Dynamics produced two major robotics news cycles last week. The larger of the two was, naturally, the electric Atlas announcement. As I write this, the sub-40 second video is steadily approaching five million views. A day prior, the company tugged at the community’s heart strings when it announced that the original hydraulic Atlas was being put out to pasture, a decade after its introduction.

The accompanying video was a celebration of the older Atlas’ journey from DARPA research project to an impressively nimble bipedal ’bot. A minute in, however, the tone shifts. Ultimately, “Farewell to Atlas” is as much a celebration as it is a blooper reel. It’s a welcome reminder that for every time the robot sticks the landing on video there are dozens of slips, falls and sputters.

Boston Dynamics' Atlas in action
Image Credits: Boston Dynamics

I’ve long championed this sort of transparency. It’s the sort of thing I would like to see more from the robotics world. Simply showcasing the highlight reel does a disservice to the effort that went into getting those shots. In many cases, we’re talking years of trial and error spent getting robots to look good on camera. When you only share the positive outcomes, you’re setting unrealistic expectations. Bipedal robots fall over. In that respect, at least, they’re just like us. As Agility put it recently, “Everyone falls sometimes, it’s how we get back up that defines us.” I would take that a step further, adding that learning how to fall well is equally important.

The company’s newly appointed CTO, Pras Velagapudi, recently told me that seeing robots fall on the job at this stage is actually a good thing. “When a robot is actually out in the world doing real things, unexpected things are going to happen,” he notes. “You’re going to see some falls, but that’s part of learning to run a really long time in real-world environments. It’s expected, and it’s a sign that you’re not staging things.”

A quick scan of Harvard’s rules for falling without injury reflects what we intuitively understand about falling as humans:

Protect your headUse your weight to direct your fallBend your kneesAvoid taking other people with you

As for robots, this IEEE Spectrum piece from last year is a great place to start.

“We’re not afraid of a fall—we’re not treating the robots like they’re going to break all the time,” Boston Dynamics CTO Aaron Saunders told the publication last year. “Our robot falls a lot, and one of the things we decided a long time ago [is] that we needed to build robots that can fall without breaking. If you can go through that cycle of pushing your robot to failure, studying the failure, and fixing it, you can make progress to where it’s not falling. But if you build a machine or a control system or a culture around never falling, then you’ll never learn what you need to learn to make your robot not fall. We celebrate falls, even the falls that break the robot.”

Image Credits: Boston Dynamics

The subject of falling also came up when I spoke with Boston Dynamics CEO Robert Playter ahead of the electric Atlas’ launch. Notably, the short video begins with the robot in a prone position. The way the robot’s legs arc around is quite novel, allowing the system to stand up from a completely flat position. At first glance, it almost feels as though the company is showing off, using the flashy move simply as a method to showcase the extremely robust custom-built actuators.

“There will be very practical uses for that,” Playter told me. “Robots are going to fall. You’d better be able to get up from prone.” He adds that the ability to get up from a prone position may also be useful for charging purposes.

Much of Boston Dynamics’ learnings around falling came from Spot. While there’s generally more stability in the quadrupedal form factor (as evidenced from decades trying and failing to kick the robots over in videos), there are simply way more hours of Spot robots working in real-world conditions.

Image Credits: Agility Robotics

“Spot’s walking something like 70,000 kms a year on factory floors, doing about 100,000 inspections per month,” adds Playter. “They do fall, eventually. You have to be able to get back up. Hopefully you get your fall rate down — we have. I think we’re falling once every 100-200 kms. The fall rate has really gotten small, but it does happen.”

Playter adds that the company has a long history of being “rough” on its robots. “They fall, and they’ve got to be able to survive. Fingers can’t fall off.”

Watching the above Atlas outtakes, it’s hard not to project a bit of human empathy onto the ’bot. It really does appear to fall like a human, drawing its extremities as close to its body as possible, to protect them from further injury.

When Agility added arms to Digit, back in 2019, it discussed the role they play in falling. “For us, arms are simultaneously a tool for moving through the world — think getting up after a fall, waving your arms for balance, or pushing open a door — while also being useful for manipulating or carrying objects,” co-founder Jonathan Hurst noted at the time.

I spoke a bit to Agility about the topic at Modex earlier this year. Video of a Digit robot falling over on a convention floor a year prior had made the social media rounds. “With a 99% success rate over about 20 hours of live demos, Digit still took a couple of falls at ProMat,” Agility noted at the time. “We have no proof, but we think our sales team orchestrated it so they could talk about Digits quick-change limbs and durability.”

As with the Atlas video, the company told me that something akin to a fetal position is useful in terms of protecting the robot’s legs and arms.

The company has been using reinforcement learning to help fallen robots right themselves. Agility shut off Digit’s obstacle avoidance for the above video to force a fall. In the video, the robot uses its arms to mitigate the fall as much as possible. It then utilizes its reinforcement learnings to return to a familiar position from which it is capable of standing again with a robotic pushup.

One of humanoid robots’ main selling points is their ability to slot into existing workflows — these factories and warehouses are known as “brownfield,” meaning they weren’t custom built for automation. In many existing cases of factory automation, errors mean the system effectively shuts down until a human intervenes.

“Rescuing a humanoid robot is not going to be trivial,” says Playter, noting that these systems are heavy and can be difficult to manually right. “How are you going to do that if it can’t get itself off the ground?”

If these systems are truly going to ensure uninterrupted automation, they’ll need to fall well and get right back up again.

“Every time Digit falls, we learn something new,” adds Velagapudi. “When it comes to bipedal robotics, falling is a wonderful teacher.”

Digital check mark over data storage signifying quailty checking.

Anomalo's machine learning approach to data quality is growing like gangbusters

Digital check mark over data storage signifying quailty checking.

Image Credits: da-kuk / Getty Images

When Anomalo’s co-founders left Instacart in 2018, they thought they could put machine learning to work to solve data-quality problems inherent in large datasets. Five years later, the company’s idea is even more relevant as data quality takes center stage with large language models.

Today, the startup announced a $33 million Series B, equaling their 2021 Series A and bringing the total raised to $72 million, according to the company.

Co-founder and CEO Elliot Shmukler says that the thesis they started with has become validated over the years. “If you’re going to use data to do anything, whether it’s dashboarding, decision-making, or these days to power generative AI applications, then you need [a tool] that’s actually monitoring that data and making sure it’s correct, of high quality and ready to be used,” Shmukler told TechCrunch.

As companies store increasingly large amounts of data in cloud storage and data warehouses like Databricks and Snowflake, this need has only become more pronounced, he says. But in a time when everyone is looking to cut costs, they came up with a way to limit the data that Anomalo monitors to certain datasets, instead of monitoring everything, to help lower customer bills. “You can reserve our kind of full ML and AI [solution] for the tables and datasets that really need it,” he said.

The approach is working. Shmukler indicated that the company has grown 15x since that Series A when he told TechCrunch the revenue was around $1 million. That would make today’s revenue close to $15 million. What’s more, for the company’s recent fiscal third quarter, he reported that annual recurring revenue grew a whopping 177%, growth numbers we haven’t seen in some time from early-stage enterprise startups.

Shmukler says he still understands that while investors obviously welcome this kind of growth, they are still very much in efficiency mode, and as CEO he has to find ways to walk that line. “Investors still love high growth, they just don’t want you to light the cash on fire,” he said.

To help find the proper balance, the company has set a couple of goals to find that sweet spot between growth and efficiency. “Our growth goal was based on percentage growth in ARR, and our efficiency is actually based on burn multiple, which is emerging as one of these efficiency metrics that investors are paying attention to. And so we see that efficiency metric of burn multiple as a kind of counterbalance on our growth,” he said.

As the company’s revenue grows, they have been hiring and are currently up to 50 employees with plans to double that with the new money. The company told us in 2021, when it had less than 10 people, that it saw diversity as one of its core goals. Shmukler says it’s a work in progress, but of the seven executives they hired since the A round, four are women. He says a third of the engineering group are women, and they are working to close the gap in that number. He believes that having women in leadership roles will help attract others.

“Having women leaders in place and women engineering managers in place really has been tremendous in terms of attracting women candidates for all of our roles. And so I think that’s going to serve us well as we double the company again,” he said.

SignalFire led the $33 million Series B investment with participation from strategic investor Databricks Ventures. Previous investors Norwest Venture Partners, Two Sigma Ventures and Foundation Capital also participated in the round. It’s particularly interesting that the company has attracted the attention of one of the leading data analytics startups in Databricks, which had a $43 billion valuation as of last September, when it raised another $500 million.

CodeSignal screen

CodeSignal launches a learning platform with an AI-powered guide

CodeSignal screen

Image Credits: CodeSignal

Technical assessment company CodeSignal, which counts Index Ventures and Menlo Ventures as backers, is launching a learning platform called CodeSignal Learn. The new platform, aimed at audiences wanting to learn technical subjects, also has an AI-powered bot called Cosmo, which helps users with learning.

The company said that there are hundreds of courses available at launch ranging from introduction to programming, tutorials on specific languages, data analytics and machine learning. Eventually, CodeSignal wants to expand into some non-technical areas — such as management skills — that accelerate career paths.

How does the platform work?

CodeSignal Learn is currently accepting users through a waitlist. Once you get in, the Cosmo bot asks you questions like what you want to learn and what is your skill level. Based on that, it forms a course path for you.

Image Credits: CodeSignal

The platform has two tiers, and the company has chosen a gamified path to monetize its product. With the free tier, you can go through lessons on your own. But as soon as you ask a question to the Cosmo bot, or ask it to help you with an evaluation of your code, you lose one energy bar. Free users get five energy bars with a bar recharging every four hours. This whole process has a mobile game-like feel to it.

If you don’t want to go through any of this hassle, you can pay up to $24.99 per month and get unlimited energy.

Users can learn about different subjects through text or convert the lessons into a slideshow or video-styled presentation. The company said that it has created courses with subject matter experts to put emphasis on practice-first learning, so the courses usually have 90% of content that requires users to work with an integrated development environment (IDE).

Why CodeSignal is developing a learning platform?

The company was originally named CodeFight when it was founded in 2014 as a competitive coding platform. But as it evolved, the startup focused more on technical assessment and rebranded as CodeSignal in 2018.

CodeSignal CEO Tigran Sloyan told TechCrunch over a call that the co-founders started the company with a mission to discover skilled workers and develop the skills that will shape the future. Sloyan said that the startup wanted to do it all and make an educational product as well.

“We quickly realized that focus is everything. So we thought we had to fix assessments first as they are fundamental for both hiring and development,” he said. With the new platform, the company is targeting the idea of skill development for people again.

“We felt that until we figured out a great one-on-one tutoring product we couldn’t have a revolutionary educational product. But with the rise in AI-based tech, we felt this is the right moment,” Sloyan said.

While the company thinks that the learning platform can upskill a wide audience, CodeSignal thought about its customers and how it could work on closing the skill gap for their workers or folks they want to hire.

Competition and the road ahead

Multiple companies are competing in the market of both technical assessments and skill building. Microsoft-owned LinkedIn has both learning and assessment modules. Last year, it launched an AI-powered assistant for learning. Pluralsight, which Sloyan feels is CodeSignal’s biggest competitor, also has an assessment product. In the learning space, there are a ton of competitors such as Coursera and Udemy, along with more focused ones like Guild and Articulate.

Sloyan thinks that CodeSignal has an edge because it has a practice-first approach to learning. Plus, he believes that the Cosmo AI bot is akin to having a private tutor by your side.

The company is also launching CodeSignal for teams and beta testing the enterprise products, which will offer a more customized version to large organizations. The CEO expects that the learning product will generate 50% of the company’s revenue in two years.

CodeSignal raised its last investment with a Series C raise of $50 million in 2021. While the startup is not profitable yet, it is not looking to raise another round anytime soon. The company has raised $87.5 million to date. The company employs 200 people, almost three times more than the 70 people it had on payroll in 2021.

Founders of GoStudent Gregor Müller & Felix Ohswald

GoStudent, the online learning platform, says it's now profitable

Founders of GoStudent Gregor Müller & Felix Ohswald

Image Credits: GoStudent

GoStudent — the online tutoring marketplace last valued at $3.2 billion — has carved out a position for itself as one of the biggest and more popular startups to come out of Vienna, Austria, with 11 million families and 23,000 tutors on the platform. Now it’s adding another distinction to the list: It is profitable. The company’s CEO Felix Ohswald told TechCrunch that the company is now in the black across its global footprint.

“And not just EBITDA profitable,” said Ohswald, who co-founded the company with COO Gregor Müller in 2016. The company is now both EBITDA positive and has positive operating cashflow across GoStudent and the GoStudent Group.

Alongside online trading platform Bitpanda and the fitness tracker Runtastic, acquired by Adidas, eight-year-old GoStudent is one of the very few Austrian startups that has made a name for itself in the international tech scene.

Its arrival at profitability has not been without bumps, however.

Amid a boom for online learning (and other technology favoring remote and virtual experiences) — the market crashed for it and many other technology companies. Not only did cash burn become a major issue after years of companies going hell for leather on scaling, but for many customer demand started to sharply drop off, creating a critical situation. At GoStudent, the company lost €89 million in 2021, with that figure ballooning to €220 million in 2022. That led to several rounds of layoffs, and major cost cuts.

Back from the heights

GoStudent’s turnaround comes after a phase that Ohswald himself referred to as “crazy” hyperscaling. “From 2019 to 2022, we scaled our core business model to more than €100 million in revenue and it was amazing, crazy growth from zero basically within two years.” But, he added, the company also had “a cash burn of over €150 million in 2022 alone.”

As markets turned, and despite having raised hundreds millions of euros in funding, the company knew it couldn’t continue on that path. In 2023, it reduced its burn rate by 70%, but it still wasn’t enough. In a LinkedIn post only a few months ago, in January of this year, Ohswald confirmed that the company was conducting another round of layoffs — the third since 2022.

These restructurings were “tough moments,” Müller said, but the company had to figure out how to keep on growing without spending as much. “At least we learned a lot now. We have a better idea of how and where to scale, of the key things we need to nail and be more careful about.”

GoStudent’s hypergrowth wasn’t only hubris. If the digital transformation brought along by Covid-19 lifted many boats, that was particularly true for edtech, and even more so for GoStudent. The company went from having to convince parents of the merits of online tutoring to becoming the go-to solution for schoolchildren in need of educational help.

But even if that sentiment has continued to carry business through for the company, it was still burning cash and needed to break the circuit. The first cuts that the company made outside of its core business were easy. No more lavish parties. A pause on acquisitions. No attempts to expand to markets where online tutoring wasn’t already established, such as Sweden, or where it had to lower its prices too much to compete, like Latin America.

Others were more painful, like exiting the U.S. after only a few months and high-profile entry to take on that very big and mure, but also very competitive, market.

GoStudent no longer aims to be present in 20 countries. It is instead refocusing on Europe. And outside of German-speaking countries, it will “adopt a more organic growth strategy.” Quite tellingly, when its former chief growth officer and early employee Laura Warnier exited the company, she was replaced by a chief marketing officer, former Delivery Hero staffer Dan Zbijowski. Goodbye top-line growth, hello bottom-line growth.

A long way to the top

Spending less while still growing into its big dreams will be a fine balancing act for GoStudent, which still has a stated mission to “build the #1 global school and unlock the potential of every student through personalized tutoring.” Scaling at lower costs inevitably will butt up against the kind of personalized approach that its customers might demand.

GoStudent is not a school right now: Its offering still very much falls under tutoring, not teaching. But its acquisition of Studienkreis in 2022 does speak to how it will be evolving that. Founded in 1974, Studienkreis has 1,000 physical learning centers, mostly in Germany, and GoStudent is now using them to double down on hybrid learning.

GoStudent uses its warchest to acquire large network of traditional tutoring centres in Europe

“We believe that the future of education is hybrid,” Ohswald said. “Glocal” was another keyword he used; while education curricula are national or even regional, GoStudent can add value by leveraging technology to make sure each kid finds the right tutor, independent of location.

Most GoStudent’s tutors are university students, Müller said, and this younger demographic makes it easier for them to click with pupils while serving as role models, too.

According to Ohswald, that’s a reinforcement that many kids these days need, as they have to live under social media pressure that didn’t exist when we grew up. “Having this moment where a person sits down with you individually builds some confidence is often worth much more than improving your grades.”

In the GoStudent Future of Education Report 2024, based on answers from more than 5,000 parents, the company found that families are looking for a more personalized approach to their child’s learning. “Obviously, grade improvement is one key thing that the parents are looking for; and if they don’t see that, they are not satisfied.” But they also want their kids to get better at problem solving and other life skills.

GoStudent can come in where schools fall short, but such a far-ranging mission requires outstanding teachers. Presumably, pay that keeps up with inflation would help attracting and retaining these. Especially in light of a recent petition from some of its 23,000 tutors complaining that they “receive less than 50% of parents’ fees while undertaking almost 100% of the preparation and administrative work.” However, GoStudent’s founders see this differently.

When asked about pay increases, Ohswald went into a spiel on how purpose-driven tutors feel fulfilled from seeing students succeed. But perhaps more tangibly, GoStudent is working on leveraging AI to make efficiency improvements on “things that otherwise would take a lot of time so the teacher can focus on teaching and not on grading work,” Müller said.

A fine balancing act

GoStudent has three priorities for 2024, Ohswald told TechCrunch: Remaining cash-flow positive, staying true to its goal of putting students first and showing how AI allows GoStudent to scale its business in a capital-efficient way.

The key here, Ohswald said, is for GoStudent to demonstrate “how we leverage AI so that we can scale operations 100 times without the need of hundreds more people. AI allows us to recruit teachers in a much more automated way, help teachers better teach their [students] and help our support and operations people on the ground hyperscale [this] without spending money on it.”

M&A is another thing GoStudent won’t spend money on “anytime soon,” but the founders are glad they did. “I really believe having been in the position to do those acquisitions that fast in such a market environment where raising capital was easier will help us massively,” Müller said.

GoStudent is already seeing the value of becoming a group, and not just a company. Tus Medias, a network of tutor marketplaces, is proving to be a solid customer acquisition channel, but also an alternative that tutors and parents can turn to should they be unhappy with what GoStudent offers.

A recurring complaint from parents is that GoStudent pushes them to multi-year commitments, only to make cancellations difficult. GoStudent retorts that education requires consistency, not a one-time fix. But of course, contracts are also better at providing GoStudent with steady revenue. In fairness, it also makes it more likely for its tutors to get a relatively stable volume of work; and more than they would if they had to find clients on their own.

Still, unhappy parents regularly take issues with GoStudent to the press. One heavily relayed case took place in the U.K. in 2021, when a father found out that his daughter’s GoStudent tutor was barred from teaching. The company apologized and said it had already changed its hiring practices, which include background checks (Enhanced DBA, in the U.K.) and a code of conduct that forbids teachers from contacting students under 16 using WhatsApp “under any circumstances.”

Child safety is one reason why the company would invest in building its own tools, such as GoChat. Sure, it could keep on using external solutions. After all, it spent its first three years as a WhatsApp homework chat. But in-house solutions make it easier to prevent tutors from getting their students’ phone numbers and to track what’s happening during the class.

GoStudent also finally abandoned Zoom in favor of its own online classroom, GoClass, based on previous developments from Tus Medias. There may be bugs for now, but it’s also a reminder that GoStudent doesn’t only want to use ready-made tools: It wants to come up with technological innovations to teach better. For instance, one recent addition is GoVR, a virtual reality platform for language learning.

GoStudent adds another $95M to its war chest to go after VR and AI-enhanced tutoring

All the talk about AI, VR and hybrid learning may have been helpful in raising GoStudent’s latest funding, a $95 million mix of equity and debt that it secured in August.

But more than anything, it is profitability that opens checkbooks and give companies more options. It gives GoStudent the option to raise more debt to avoid more dilution, pick a different structure or just not raise additional capital. It’s in our hands to find the right strategy,” Ohswald said. That’s true on the financing front, but also elsewhere.

Boston Dynamics... robot down!

Humanoid robots are learning to fall well

Boston Dynamics... robot down!

Image Credits: Boston Dynamics

The savvy marketers at Boston Dynamics produced two major robotics news cycles last week. The larger of the two was, naturally, the electric Atlas announcement. As I write this, the sub-40 second video is steadily approaching five million views. A day prior, the company tugged at the community’s heart strings when it announced that the original hydraulic Atlas was being put out to pasture, a decade after its introduction.

The accompanying video was a celebration of the older Atlas’ journey from DARPA research project to an impressively nimble bipedal ’bot. A minute in, however, the tone shifts. Ultimately, “Farewell to Atlas” is as much a celebration as it is a blooper reel. It’s a welcome reminder that for every time the robot sticks the landing on video there are dozens of slips, falls and sputters.

Boston Dynamics' Atlas in action
Image Credits: Boston Dynamics

I’ve long championed this sort of transparency. It’s the sort of thing I would like to see more from the robotics world. Simply showcasing the highlight reel does a disservice to the effort that went into getting those shots. In many cases, we’re talking years of trial and error spent getting robots to look good on camera. When you only share the positive outcomes, you’re setting unrealistic expectations. Bipedal robots fall over. In that respect, at least, they’re just like us. As Agility put it recently, “Everyone falls sometimes, it’s how we get back up that defines us.” I would take that a step further, adding that learning how to fall well is equally important.

The company’s newly appointed CTO, Pras Velagapudi, recently told me that seeing robots fall on the job at this stage is actually a good thing. “When a robot is actually out in the world doing real things, unexpected things are going to happen,” he notes. “You’re going to see some falls, but that’s part of learning to run a really long time in real-world environments. It’s expected, and it’s a sign that you’re not staging things.”

A quick scan of Harvard’s rules for falling without injury reflects what we intuitively understand about falling as humans:

Protect your headUse your weight to direct your fallBend your kneesAvoid taking other people with you

As for robots, this IEEE Spectrum piece from last year is a great place to start.

“We’re not afraid of a fall—we’re not treating the robots like they’re going to break all the time,” Boston Dynamics CTO Aaron Saunders told the publication last year. “Our robot falls a lot, and one of the things we decided a long time ago [is] that we needed to build robots that can fall without breaking. If you can go through that cycle of pushing your robot to failure, studying the failure, and fixing it, you can make progress to where it’s not falling. But if you build a machine or a control system or a culture around never falling, then you’ll never learn what you need to learn to make your robot not fall. We celebrate falls, even the falls that break the robot.”

Image Credits: Boston Dynamics

The subject of falling also came up when I spoke with Boston Dynamics CEO Robert Playter ahead of the electric Atlas’ launch. Notably, the short video begins with the robot in a prone position. The way the robot’s legs arc around is quite novel, allowing the system to stand up from a completely flat position. At first glance, it almost feels as though the company is showing off, using the flashy move simply as a method to showcase the extremely robust custom-built actuators.

“There will be very practical uses for that,” Playter told me. “Robots are going to fall. You’d better be able to get up from prone.” He adds that the ability to get up from a prone position may also be useful for charging purposes.

Much of Boston Dynamics’ learnings around falling came from Spot. While there’s generally more stability in the quadrupedal form factor (as evidenced from decades trying and failing to kick the robots over in videos), there are simply way more hours of Spot robots working in real-world conditions.

Image Credits: Agility Robotics

“Spot’s walking something like 70,000 kms a year on factory floors, doing about 100,000 inspections per month,” adds Playter. “They do fall, eventually. You have to be able to get back up. Hopefully you get your fall rate down — we have. I think we’re falling once every 100-200 kms. The fall rate has really gotten small, but it does happen.”

Playter adds that the company has a long history of being “rough” on its robots. “They fall, and they’ve got to be able to survive. Fingers can’t fall off.”

Watching the above Atlas outtakes, it’s hard not to project a bit of human empathy onto the ’bot. It really does appear to fall like a human, drawing its extremities as close to its body as possible, to protect them from further injury.

When Agility added arms to Digit, back in 2019, it discussed the role they play in falling. “For us, arms are simultaneously a tool for moving through the world — think getting up after a fall, waving your arms for balance, or pushing open a door — while also being useful for manipulating or carrying objects,” co-founder Jonathan Hurst noted at the time.

I spoke a bit to Agility about the topic at Modex earlier this year. Video of a Digit robot falling over on a convention floor a year prior had made the social media rounds. “With a 99% success rate over about 20 hours of live demos, Digit still took a couple of falls at ProMat,” Agility noted at the time. “We have no proof, but we think our sales team orchestrated it so they could talk about Digits quick-change limbs and durability.”

As with the Atlas video, the company told me that something akin to a fetal position is useful in terms of protecting the robot’s legs and arms.

The company has been using reinforcement learning to help fallen robots right themselves. Agility shut off Digit’s obstacle avoidance for the above video to force a fall. In the video, the robot uses its arms to mitigate the fall as much as possible. It then utilizes its reinforcement learnings to return to a familiar position from which it is capable of standing again with a robotic pushup.

One of humanoid robots’ main selling points is their ability to slot into existing workflows — these factories and warehouses are known as “brownfield,” meaning they weren’t custom built for automation. In many existing cases of factory automation, errors mean the system effectively shuts down until a human intervenes.

“Rescuing a humanoid robot is not going to be trivial,” says Playter, noting that these systems are heavy and can be difficult to manually right. “How are you going to do that if it can’t get itself off the ground?”

If these systems are truly going to ensure uninterrupted automation, they’ll need to fall well and get right back up again.

“Every time Digit falls, we learn something new,” adds Velagapudi. “When it comes to bipedal robotics, falling is a wonderful teacher.”