Tag Archives: runs

#437269 DeepMind’s Newest AI Programs Itself ...

Posted on July 27, 2020 by Android

When Deep Blue defeated world chess champion Garry Kasparov in 1997, it may have seemed artificial intelligence had finally arrived. A computer had just taken down one of the top chess players of all time. But it wasn’t to be.

Though Deep Blue was meticulously programmed top-to-bottom to play chess, the approach was too labor-intensive, too dependent on clear rules and bounded possibilities to succeed at more complex games, let alone in the real world. The next revolution would take a decade and a half, when vastly more computing power and data revived machine learning, an old idea in artificial intelligence just waiting for the world to catch up.

Today, machine learning dominates, mostly by way of a family of algorithms called deep learning, while symbolic AI, the dominant approach in Deep Blue’s day, has faded into the background.

Key to deep learning’s success is the fact the algorithms basically write themselves. Given some high-level programming and a dataset, they learn from experience. No engineer anticipates every possibility in code. The algorithms just figure it.

Now, Alphabet’s DeepMind is taking this automation further by developing deep learning algorithms that can handle programming tasks which have been, to date, the sole domain of the world’s top computer scientists (and take them years to write).

In a paper recently published on the pre-print server arXiv, a database for research papers that haven’t been peer reviewed yet, the DeepMind team described a new deep reinforcement learning algorithm that was able to discover its own value function—a critical programming rule in deep reinforcement learning—from scratch.

Surprisingly, the algorithm was also effective beyond the simple environments it trained in, going on to play Atari games—a different, more complicated task—at a level that was, at times, competitive with human-designed algorithms and achieving superhuman levels of play in 14 games.

DeepMind says the approach could accelerate the development of reinforcement learning algorithms and even lead to a shift in focus, where instead of spending years writing the algorithms themselves, researchers work to perfect the environments in which they train.

Pavlov’s Digital Dog
First, a little background.

Three main deep learning approaches are supervised, unsupervised, and reinforcement learning.

The first two consume huge amounts of data (like images or articles), look for patterns in the data, and use those patterns to inform actions (like identifying an image of a cat). To us, this is a pretty alien way to learn about the world. Not only would it be mind-numbingly dull to review millions of cat images, it’d take us years or more to do what these programs do in hours or days. And of course, we can learn what a cat looks like from just a few examples. So why bother?

While supervised and unsupervised deep learning emphasize the machine in machine learning, reinforcement learning is a bit more biological. It actually is the way we learn. Confronted with several possible actions, we predict which will be most rewarding based on experience—weighing the pleasure of eating a chocolate chip cookie against avoiding a cavity and trip to the dentist.

In deep reinforcement learning, algorithms go through a similar process as they take action. In the Atari game Breakout, for instance, a player guides a paddle to bounce a ball at a ceiling of bricks, trying to break as many as possible. When playing Breakout, should an algorithm move the paddle left or right? To decide, it runs a projection—this is the value function—of which direction will maximize the total points, or rewards, it can earn.

Move by move, game by game, an algorithm combines experience and value function to learn which actions bring greater rewards and improves its play, until eventually, it becomes an uncanny Breakout player.

Learning to Learn (Very Meta)
So, a key to deep reinforcement learning is developing a good value function. And that’s difficult. According to the DeepMind team, it takes years of manual research to write the rules guiding algorithmic actions—which is why automating the process is so alluring. Their new Learned Policy Gradient (LPG) algorithm makes solid progress in that direction.

LPG trained in a number of toy environments. Most of these were “gridworlds”—literally two-dimensional grids with objects in some squares. The AI moves square to square and earns points or punishments as it encounters objects. The grids vary in size, and the distribution of objects is either set or random. The training environments offer opportunities to learn fundamental lessons for reinforcement learning algorithms.

Only in LPG’s case, it had no value function to guide that learning.

Instead, LPG has what DeepMind calls a “meta-learner.” You might think of this as an algorithm within an algorithm that, by interacting with its environment, discovers both “what to predict,” thereby forming its version of a value function, and “how to learn from it,” applying its newly discovered value function to each decision it makes in the future.

Prior work in the area has had some success, but according to DeepMind, LPG is the first algorithm to discover reinforcement learning rules from scratch and to generalize beyond training. The latter was particularly surprising because Atari games are so different from the simple worlds LPG trained in—that is, it had never seen anything like an Atari game.

Time to Hand Over the Reins? Not Just Yet
LPG is still behind advanced human-designed algorithms, the researchers said. But it outperformed a human-designed benchmark in training and even some Atari games, which suggests it isn’t strictly worse, just that it specializes in some environments.

This is where there’s room for improvement and more research.

The more environments LPG saw, the more it could successfully generalize. Intriguingly, the researchers speculate that with enough well-designed training environments, the approach might yield a general-purpose reinforcement learning algorithm.

At the least, though, they say further automation of algorithm discovery—that is, algorithms learning to learn—will accelerate the field. In the near term, it can help researchers more quickly develop hand-designed algorithms. Further out, as self-discovered algorithms like LPG improve, engineers may shift from manually developing the algorithms themselves to building the environments where they learn.

Deep learning long ago left Deep Blue in the dust at games. Perhaps algorithms learning to learn will be a winning strategy in the real world too.

Image credit: Mike Szczepanski / Unsplash Continue reading →

Posted in Human Robots

#437209 A Renaissance of Genomics and Drugs Is ...

Posted on June 26, 2020 by Android

The causes of aging are extremely complex and unclear. But with longevity clinical trials increasing, more answers—and questions—are emerging than ever before.

With the dramatic demonetization of genome reading and editing over the past decade, and Big Pharma, startups, and the FDA starting to face aging as a disease, we are starting to turn those answers into practical ways to extend our healthspan.

In this article, I’ll explore how genome sequencing and editing, along with new classes of anti-aging drugs, are augmenting our biology to further extend our healthy lives.

Genome Sequencing and Editing
Your genome is the software that runs your body. A sequence of 3.2 billion letters makes you “you.” These base pairs of A’s, T’s, C’s, and G’s determine your hair color, your height, your personality, your propensity for disease, your lifespan, and so on.

Until recently, it’s been very difficult to rapidly and cheaply “read” these letters—and even more difficult to understand what they mean. Since 2001, the cost to sequence a whole human genome has plummeted exponentially, outpacing Moore’s Law threefold. From an initial cost of $3.7 billion, it dropped to $10 million in 2006, and to $1,500 in 2015.

Today, the cost of genome sequencing has dropped below $600, and according to Illumina, the world’s leading sequencing company, the process will soon cost about $100 and take about an hour to complete.

This represents one of the most powerful and transformative technology revolutions in healthcare. When we understand your genome, we’ll be able to understand how to optimize “you.”

We’ll know the perfect foods, the perfect drugs, the perfect exercise regimen, and the perfect supplements, just for you.
We’ll understand what microbiome types, or gut flora, are ideal for you (more on this in a later article).
We’ll accurately predict how specific sedatives and medicines will impact you.
We’ll learn which diseases and illnesses you’re most likely to develop and, more importantly, how to best prevent them from developing in the first place (rather than trying to cure them after the fact).

CRISPR Gene Editing
In addition to reading the human genome, scientists can now edit a genome using a naturally occurring biological system discovered in 1987 called CRISPR/Cas9.

Short for Clustered Regularly Interspaced Short Palindromic Repeats and CRISPR-associated protein 9, the editing system was adapted from a naturally-occurring defense system found in bacteria.

Here’s how it works. The bacteria capture snippets of DNA from invading viruses (or bacteriophage) and use them to create DNA segments known as CRISPR arrays. The CRISPR arrays allow the bacteria to “remember” the viruses (or closely related ones), and defend against future invasions. If the viruses attack again, the bacteria produce RNA segments from the CRISPR arrays to target the viruses’ DNA. The bacteria then use Cas9 to cut the DNA apart, which disables the virus.

Most importantly, CRISPR is cheap, quick, easy to use, and more accurate than all previous gene editing methods. As a result, CRISPR/Cas9 has swept through labs around the world as the way to edit a genome. A short search in the literature will show an exponential rise in the number of CRISPR-related publications and patents.

2018: Filled With CRISPR Breakthroughs
Early results are impressive. Researchers have used CRISPR to genetically engineer cocaine resistance into mice, reverse the gene defect causing Duchenne muscular dystrophy (DMD) in dogs, and reduce genetic deafness in mice.

Already this year, CRISPR-edited immune cells have been shown to successfully kill cancer cells in human patients. Researchers have discovered ways to activate CRISPR with light and use the gene-editing technology to better understand Alzheimer’s disease progression.

With great power comes great responsibility, and the opportunity for moral and ethical dilemmas. In 2015, Chinese scientists sparked global controversy when they first edited human embryo cells in the lab with the goal of modifying genes that would make the child resistant to smallpox, HIV, and cholera. Three years later, in November 2018, researcher He Jiankui informed the world that the first set of CRISPR-engineered female twins had been delivered.

To accomplish his goal, Jiankui deleted a region of a receptor on the surface of white blood cells known as CCR5, introducing a rare, natural genetic variation that makes it more difficult for HIV to infect its favorite target, white blood cells. Because Jiankui forged ethical review documents and misled doctors in the process, he was sentenced to three years in prison and fined $429,000 last December.

Coupled with significant ethical conversations necessary for progress, CRISPR will soon provide us the tools to eliminate diseases, create hardier offspring, produce new environmentally resistant crops, and even wipe out pathogens.

Senolytics, Nutraceuticals, and Pharmaceuticals
Over the arc of your life, the cells in your body divide until they reach what is known as the Hayflick limit, or the number of times a normal human cell population will divide before cell division stops, which is typically about 50 divisions.

What normally follows next is programmed cell death or destruction by the immune system. A very small fraction of cells, however, become senescent cells and evade this fate to linger indefinitely. These lingering cells secrete a potent mix of molecules that triggers chronic inflammation, damages the surrounding tissue structures, and changes the behavior of nearby cells for the worse. Senescent cells appear to be one of the root causes of aging, causing everything from fibrosis and blood vessel calcification to localized inflammatory conditions such as osteoarthritis to diminished lung function.

Fortunately, both the scientific and entrepreneurial communities have begun to work on senolytic therapies, moving the technology for selectively destroying senescent cells out of the laboratory and into a half-dozen startup companies.

Prominent companies in the field include the following:

Unity Biotechnology is developing senolytic medicines to selectively eliminate senescent cells with an initial focus on delivering localized therapy in osteoarthritis, ophthalmology, and pulmonary disease.

Oisin Biotechnologies is pioneering a programmable gene therapy that can destroy cells based on their internal biochemistry.

SIWA Therapeutics is working on an immunotherapy approach to the problem of senescent cells.

In recent years, researchers have identified or designed a handful of senolytic compounds that can curb aging by regulating senescent cells. Two of these drugs that have gained mainstay research traction are rapamycin and metformin.

(1) Rapamycin

Originally extracted from bacteria found on Easter Island, rapamycin acts on the m-TOR (mechanistic target of rapamycin) pathway to selectively block a key protein that facilitates cell division. Currently, rapamycin derivatives are widely used for immunosuppression in organ and bone marrow transplants. Research now suggests that use results in prolonged lifespan and enhanced cognitive and immune function.

PureTech Health subsidiary resTORbio (which went public in 2018) is working on a rapamycin-based drug intended to enhance immunity and reduce infection. Their clinical-stage RTB101 drug works by inhibiting part of the mTOR pathway.

Results of the drug’s recent clinical trial include decreased incidence of infection, improved influenza vaccination response, and a 30.6 percent decrease in respiratory tract infection.

Impressive, to say the least.

(2) Metformin

Metformin is a widely-used generic drug for mitigating liver sugar production in Type 2 diabetes patients. Researchers have found that metformin also reduces oxidative stress and inflammation, which otherwise increase as we age. There is strong evidence that metformin can augment cellular regeneration and dramatically mitigate cellular senescence by reducing both oxidative stress and inflammation.

Over 100 studies registered on ClinicalTrials.gov are currently following up on strong evidence of metformin’s protective effect against cancer.

(3) Nutraceuticals and NAD+

Beyond cellular senescence, certain critical nutrients and proteins tend to decline as a function of age. Nutraceuticals combat aging by supplementing and replenishing these declining nutrient levels.

NAD+ exists in every cell, participating in every process from DNA repair to creating the energy vital for cellular processes. It’s been shown that NAD+ levels decline as we age.

The Elysium Health Basis supplement aims to elevate NAD+ levels in the body to extend one’s lifespan. Elysium’s first clinical study reports that Basis increases NAD+ levels consistently by a sustained 40 percent.

Conclusion
These are just a taste of the tremendous momentum that longevity and aging technology has right now. As artificial intelligence and quantum computing transform how we decode our DNA and how we discover drugs, genetics and pharmaceuticals will become truly personalized.

The next article in this series will demonstrate how artificial intelligence is converging with genetics and pharmaceuticals to transform how we approach longevity, aging, and vitality.

We are edging closer toward a dramatically extended healthspan—where 100 is the new 60. What will you create, where will you explore, and how will you spend your time if you are able to add an additional 40 healthy years to your life?

Join Me
(1) A360 Executive Mastermind: If you’re an exponentially and abundance-minded entrepreneur who would like coaching directly from me, consider joining my Abundance 360 Mastermind, a highly selective community of 360 CEOs and entrepreneurs who I coach for 3 days every January in Beverly Hills, Ca. Through A360, I provide my members with context and clarity about how converging exponential technologies will transform every industry. I’m committed to running A360 for the course of an ongoing 25-year journey as a “countdown to the Singularity.”

If you’d like to learn more and consider joining our 2021 membership, apply here.

(2) Abundance-Digital Online Community: I’ve also created a Digital/Online community of bold, abundance-minded entrepreneurs called Abundance-Digital. Abundance-Digital is Singularity University’s ‘onramp’ for exponential entrepreneurs—those who want to get involved and play at a higher level. Click here to learn more.

(Both A360 and Abundance-Digital are part of Singularity University—your participation opens you to a global community.)

This article originally appeared on diamandis.com. Read the original article here.

Image Credit: Arek Socha from Pixabay Continue reading →

Posted in Human Robots

#437150 AI Is Getting More Creative. But Who ...

Posted on May 26, 2020 by Android

Creativity is a trait that makes humans unique from other species. We alone have the ability to make music and art that speak to our experiences or illuminate truths about our world. But suddenly, humans’ artistic abilities have some competition—and from a decidedly non-human source.

Over the last couple years there have been some remarkable examples of art produced by deep learning algorithms. They have challenged the notion of an elusive definition of creativity and put into perspective how professionals can use artificial intelligence to enhance their abilities and produce beyond the known boundaries.

But when creativity is the result of code written by a programmer, using a format given by a software engineer, featuring private and public datasets, how do we assign ownership of AI-generated content, and particularly that of artwork? McKinsey estimates AI will annually generate value of $3.5 to $5.8 trillion across various sectors.

In 2018, a portrait that was christened Edmond de Belamy was made in a French art collective called Obvious. It used a database with 15,000 portraits from the 1300s to the 1900s to train a deep learning algorithm to produce a unique portrait. The painting sold for $432,500 in a New York auction. Similarly, a program called Aiva, trained on thousands of classical compositions, has released albums whose pieces are being used by ad agencies and movies.

The datasets used by these algorithms were different, but behind both there was a programmer who changed the brush strokes or musical notes into lines of code and a data scientist or engineer who fitted and “curated” the datasets to use for the model. There could also have been user-based input, and the output may be biased towards certain styles or unintentionally infringe on similar pieces of art. This shows that there are many collaborators with distinct roles in producing AI-generated content, and it’s important to discuss how they can protect their proprietary interests.

A perspective article published in Nature Machine Intelligence by Jason K. Eshraghian in March looks into how AI artists and the collaborators involved should assess their ownership, laying out some guiding principles that are “only applicable for as long as AI does not have legal parenthood, the way humans and corporations are accorded.”

Before looking at how collaborators can protect their interests, it’s useful to understand the basic requirements of copyright law. The artwork in question must be an “original work of authorship fixed in a tangible medium.” Given this principle, the author asked whether it’s possible for AI to exercise creativity, skill, or any other indicator of originality. The answer is still straightforward—no—or at least not yet. Currently, AI’s range of creativity doesn’t exceed the standard used by the US Copyright Office, which states that copyright law protects the “fruits of intellectual labor founded in the creative powers of the mind.”

Due to the current limitations of narrow AI, it must have some form of initial input that helps develop its ability to create. At the moment AI is a tool that can be used to produce creative work in the same way that a video camera is a tool used to film creative content. Video producers don’t need to comprehend the inner workings of their cameras; as long as their content shows creativity and originality, they have a proprietary claim over their creations.

The same concept applies to programmers developing a neural network. As long as the dataset they use as input yields an original and creative result, it will be protected by copyright law; they don’t need to understand the high-level mathematics, which in this case are often black box algorithms whose output it’s impossible to analyze.

Will robots and algorithms eventually be treated as creative sources able to own copyrights? The author pointed to the recent patent case of Warner-Lambert Co Ltd versus Generics where Lord Briggs, Justice of the Supreme Court of the UK, determined that “the court is well versed in identifying the governing mind of a corporation and, when the need arises, will no doubt be able to do the same for robots.”

In the meantime, Dr. Eshraghian suggests four guiding principles to allow artists who collaborate with AI to protect themselves.

First, programmers need to document their process through online code repositories like GitHub or BitBucket.

Second, data engineers should also document and catalog their datasets and the process they used to curate their models, indicating selectivity in their criteria as much as possible to demonstrate their involvement and creativity.

Third, in cases where user data is utilized, the engineer should “catalog all runs of the program” to distinguish the data selection process. This could be interpreted as a way of determining whether user-based input has a right to claim the copyright too.

Finally, the output should avoid infringing on others’ content through methods like reverse image searches and version control, as mentioned above.

AI-generated artwork is still a very new concept, and the ambiguous copyright laws around it give a lot of flexibility to AI artists and programmers worldwide. The guiding principles Eshraghian lays out will hopefully shed some light on the legislation we’ll eventually need for this kind of art, and start an important conversation between all the stakeholders involved.

Image Credit: Wikimedia Commons Continue reading →

Posted in Human Robots

#436774 AI Is an Energy-Guzzler. We Need to ...

Posted on February 29, 2020 by Android

There is a saying that has emerged among the tech set in recent years: AI is the new electricity. The platitude refers to the disruptive power of artificial intelligence for driving advances in everything from transportation to predicting the weather.

Of course, the computers and data centers that support AI’s complex algorithms are very much dependent on electricity. While that may seem pretty obvious, it may be surprising to learn that AI can be extremely power-hungry, especially when it comes to training the models that enable machines to recognize your face in a photo or for Alexa to understand a voice command.

The scale of the problem is difficult to measure, but there have been some attempts to put hard numbers on the environmental cost.

For instance, one paper published on the open-access repository arXiv claimed that the carbon emissions for training a basic natural language processing (NLP) model—algorithms that process and understand language-based data—are equal to the CO2 produced by the average American lifestyle over two years. A more robust model required the equivalent of about 17 years’ worth of emissions.

The authors noted that about a decade ago, NLP models could do the job on a regular commercial laptop. Today, much more sophisticated AI models use specialized hardware like graphics processing units, or GPUs, a chip technology popularized by Nvidia for gaming that also proved capable of supporting computing tasks for AI.

OpenAI, a nonprofit research organization co-founded by tech prophet and profiteer Elon Musk, said that the computing power “used in the largest AI training runs has been increasing exponentially with a 3.4-month doubling time” since 2012. That’s about the time that GPUs started making their way into AI computing systems.

Getting Smarter About AI Chip Design
While GPUs from Nvidia remain the gold standard in AI hardware today, a number of startups have emerged to challenge the company’s industry dominance. Many are building chipsets designed to work more like the human brain, an area that’s been dubbed neuromorphic computing.

One of the leading companies in this arena is Graphcore, a UK startup that has raised more than $450 million and boasts a valuation of $1.95 billion. The company’s version of the GPU is an IPU, which stands for intelligence processing unit.

To build a computer brain more akin to a human one, the big brains at Graphcore are bypassing the precise but time-consuming number-crunching typical of a conventional microprocessor with one that’s content to get by on less precise arithmetic.

The results are essentially the same, but IPUs get the job done much quicker. Graphcore claimed it was able to train the popular BERT NLP model in just 56 hours, while tripling throughput and reducing latency by 20 percent.

An article in Bloomberg compared the approach to the “human brain shifting from calculating the exact GPS coordinates of a restaurant to just remembering its name and neighborhood.”

Graphcore’s hardware architecture also features more built-in memory processing, boosting efficiency because there’s less need to send as much data back and forth between chips. That’s similar to an approach adopted by a team of researchers in Italy that recently published a paper about a new computing circuit.

The novel circuit uses a device called a memristor that can execute a mathematical function known as a regression in just one operation. The approach attempts to mimic the human brain by processing data directly within the memory.

Daniele Ielmini at Politecnico di Milano, co-author of the Science Advances paper, told Singularity Hub that the main advantage of in-memory computing is the lack of any data movement, which is the main bottleneck of conventional digital computers, as well as the parallel processing of data that enables the intimate interactions among various currents and voltages within the memory array.

Ielmini explained that in-memory computing can have a “tremendous impact on energy efficiency of AI, as it can accelerate very advanced tasks by physical computation within the memory circuit.” He added that such “radical ideas” in hardware design will be needed in order to make a quantum leap in energy efficiency and time.

It’s Not Just a Hardware Problem
The emphasis on designing more efficient chip architecture might suggest that AI’s power hunger is essentially a hardware problem. That’s not the case, Ielmini noted.

“We believe that significant progress could be made by similar breakthroughs at the algorithm and dataset levels,” he said.

He’s not the only one.

One of the key research areas at Qualcomm’s AI research lab is energy efficiency. Max Welling, vice president of Qualcomm Technology R&D division, has written about the need for more power-efficient algorithms. He has gone so far as to suggest that AI algorithms will be measured by the amount of intelligence they provide per joule.

One emerging area being studied, Welling wrote, is the use of Bayesian deep learning for deep neural networks.

It’s all pretty heady stuff and easily the subject of a PhD thesis. The main thing to understand in this context is that Bayesian deep learning is another attempt to mimic how the brain processes information by introducing random values into the neural network. A benefit of Bayesian deep learning is that it compresses and quantifies data in order to reduce the complexity of a neural network. In turn, that reduces the number of “steps” required to recognize a dog as a dog—and the energy required to get the right result.

A team at Oak Ridge National Laboratory has previously demonstrated another way to improve AI energy efficiency by converting deep learning neural networks into what’s called a spiking neural network. The researchers spiked their deep spiking neural network (DSNN) by introducing a stochastic process that adds random values like Bayesian deep learning.

The DSNN actually imitates the way neurons interact with synapses, which send signals between brain cells. Individual “spikes” in the network indicate where to perform computations, lowering energy consumption because it disregards unnecessary computations.

The system is being used by cancer researchers to scan millions of clinical reports to unearth insights on causes and treatments of the disease.

Helping battle cancer is only one of many rewards we may reap from artificial intelligence in the future, as long as the benefits of those algorithms outweigh the costs of using them.

“Making AI more energy-efficient is an overarching objective that spans the fields of algorithms, systems, architecture, circuits, and devices,” Ielmini said.

Image Credit: analogicus from Pixabay Continue reading →

Posted in Human Robots

#436263 Skydio 2 Review: This Is the Drone You ...

Posted on December 3, 2019 by Android

Let me begin this review by saying that the Skydio 2 is one of the most impressive robots that I have ever seen. Over the last decade, I’ve spent enough time around robots to have a very good sense of what kinds of things are particularly challenging for them, and to set my expectations accordingly. Those expectations include things like “unstructured environments are basically impossible” and “full autonomy is impractically expensive” and “robot videos rarely reflect reality.”

Skydio’s newest drone is an exception to all of this. It’s able to fly autonomously at speed through complex environments in challenging real-world conditions in a way that’s completely effortless and stress-free for the end user, allowing you to capture the kind of video that would be otherwise impossible, even (I’m guessing) for professional drone pilots. When you see this technology in action, it’s (almost) indistinguishable from magic.

Skydio 2 Price
To be clear, the Skydio 2 is not without compromises, and the price of $999 (on pre-order with delivery of the next batch expected in spring of 2020) requires some justification. But the week I’ve had with this drone has left me feeling like its fundamental autonomous capability is so far beyond just about anything that I’ve ever experienced that I’m questioning why I would every fly anything else ever again.

We’ve written extensively about Skydio, beginning in early 2016 when the company posted a video of a prototype drone dodging trees while following a dude on a bike. Even three years ago, Skydio’s tech was way better than anything we’d seen outside of a research lab, and in early 2018, they introduced their first consumer product, the Skydio R1. A little over a year later, Skydio has introduced the Skydio 2, which is smaller, smarter, and much more affordable. Here’s an overview video just to get you caught up:

Skydio sent me a Skydio 2 review unit last week, and while I’m reasonably experienced with drones in general, this is the first time I’ve tried a Skydio drone in person. I had a pretty good idea what to expect, and I was absolutely blown away. Like, I was giggling to myself while running through the woods as the drone zoomed around, deftly avoiding trees and keeping me in sight. Robots aren’t supposed to be this good.

A week is really not enough time to explore everything that the Skydio can do, especially Thanksgiving week in Washington, D.C. (a no-fly zone) in early winter. But I found a nearby state park in which I could legally and safely fly the drone, and I did my best to put the Skydio 2 through its paces.

Note: Throughout this review, we’ve got a bunch of GIFs to help illustrate different features of the drone. To fit them all in, these GIFs had to be heavily compressed. Underneath each GIF is a timestamped link to this YouTube video (also available at the bottom of the post), which you can click on to see the an extended cut of the original 4K 30 fps footage. And there’s a bunch of interesting extra video in there as well.

Skydio 2 Specs

Photo: Evan Ackerman/IEEE Spectrum

The Skydio 2 is primarily made out of magnesium, which (while light) is both heavier and more rigid and durable than plastic. The offset props (the back pair are above the body, and the front pair are below) are necessary to maintain the field of view of the navigation cameras.

The Skydio 2 both looks and feels like a well-designed and carefully thought-out drone. It’s solid, and a little on the heavy side as far as drones go—it’s primarily made out of magnesium, which (while light) is both heavier and more rigid and durable than plastic. The blue and black color scheme is far more attractive than you typically see with drones.

Photo: Evan Ackerman/IEEE Spectrum

To detect and avoid obstacles, the Skydio 2 uses an array of six 4K hemispherical cameras that feed data into an NVIDIA Jetson TX2 at 30 fps, with the drone processing a million points in 3D space per second to plan the safest path.

The Skydio 2 is built around an array of six hemispherical obstacle-avoidance cameras and the NVIDIA Jetson TX2 computing module that they’re connected to. This defines the placement of the gimbal, the motors and props, and the battery, since all of this stuff has to be as much as possible out of the view of the cameras in order for the drone to effectively avoid obstacles in any direction.

Without the bottom-mounted battery attached, the drone is quite flat. The offset props (the back pair are above the body, and the front pair are below) are necessary to maintain the field of view of the obstacle-avoidance cameras. These hemispherical cameras are on the end of each of the prop arms as well as above and below the body of the drone. They look awfully exposed, even though each is protected from ground contact by a little fin. You need to make sure these cameras are clean and smudge-free, and Skydio includes a cleaning cloth for this purpose. Underneath the drone there are slots for microSD cards, one for recording from the camera and a second one that the drone uses to store data. The attention to detail extends to the SD card insertion, which has a sloped channel that guides the card securely into its slot.

Once you snap the battery in, the drone goes from looking streamlined to looking a little chubby. Relative to other drones, the battery almost seems like an afterthought, like Skydio designed the drone and then remembered, “oops we have to add a battery somewhere, let’s just kludge it onto the bottom.” But again, the reason for this is to leave room inside the body for the NVIDIA TX2, while making sure that the battery stays out of view of the obstacle avoidance cameras.

The magnetic latching system for the battery is both solid and satisfying. I’m not sure why it’s necessary, strictly speaking, but I do like it, and it doesn’t seem like the battery will fly off even during the most aggressive maneuvers. Each battery includes an LED array that will display its charge level in 25 percent increments, as well as a button that you push to turn the drone on and off. Charging takes place via a USB-C port in the top of the drone, which I don’t like, because it means that the batteries can’t be charged on their own (like the Parrot Anafi’s battery), and that you can’t charge one battery while flying with another, like basically every other drone ever. A separate battery charger that will charge two at once is available from Skydio for an eyebrow-raising $129.

I appreciate that all of Skydio’s stuff (batteries, controller, and beacon) charges via USB-C, though. The included USB-C adapter with its beefy cable will output at up to 65 watts, which’ll charge a mostly depleted battery in under an hour. The drone turns itself on while charging, which seems unnecessary.

Photo: Evan Ackerman/IEEE Spectrum

The Skydio 2 is not foldable, making it not nearly as easy to transport as some other drones. But it does come with a nice case that mitigates this issue somewhat, and the drone plus two batteries end up as a passably flat package about the size of a laptop case.

The most obvious compromise that Skydio made with the Skydio 2 is that the drone is not foldable. Skydio CEO Adam Bry told us that adding folding joints to the arms of the Skydio 2 would have made calibrating all six cameras a nightmare and significantly impacted performance. This makes complete sense, of course, but it does mean that the Skydio 2 is not nearly as easy to transport as some other drones.

Photo: Evan Ackerman/IEEE Spectrum

Folded and unfolded: The Skydio 2 compared to the Parrot Anafi (upper left) and the DJI Mavic Pro (upper right).

The Skydio 2 does come with a very nice case that mitigates this issue somewhat, and the drone plus two batteries end up as a passably flat package about the size of a laptop case. Still, it’s just not as convenient to toss into a backpack as my Anafi, although the Mavic Mini might be even more portable.

Photo: Evan Ackerman/IEEE Spectrum

While the Skydio 2’s case is relatively compact, the non-foldable drone is overall a significantly larger package than the Parrot Anafi.

The design of the drone leads to some other compromises as well. Since landing gear would, I assume, occlude the camera system, the drone lands directly on the bottom of its battery pack, which has a slightly rubberized pad about the size of a playing card. This does’t feel particularly stable unless you end up on a very flat surface, and made me concerned for the exposed cameras underneath the drone as well as the lower set of props. I’d recommend hand takeoffs and landings—more on those later.

Skydio 2 Camera System

Photo: Evan Ackerman/IEEE Spectrum

The Skydio 2’s primary camera is a Sony IMX577 1/2.3″ 12.3-megapixel CMOS sensor. It’s mounted to a three-axis gimbal and records 4K video at 60 fps, or 1080p video at 120 fps.

The Skydio 2 comes with a three-axis gimbal supporting a 12-megapixel camera, just enough to record 4K video at 60 fps, or 1080p video at 120 fps. Skydio has provided plenty of evidence that its imaging system is at least as good if not better than other drone cameras. Tested against my Mavic Pro and Parrot Anafi, I found no reason to doubt that. To be clear, I didn’t do exhaustive pixel-peeping comparisons between them, you’re just getting my subjective opinion that the Skydio 2 has a totally decent camera that you won’t be disappointed with. I will say that I found the HDR photo function to be not all that great under the few situations in which I tested it—after looking at a few muddy sunset shots, I turned it off and was much happier.

Photo: Evan Ackerman/IEEE Spectrum

The Skydio 2’s 12-megapixel camera is solid, although we weren’t impressed with the HDR option.

The video stabilization is fantastic, to the point where watching the video footage can be underwhelming because it doesn’t reflect the motion of the drone. I almost wish there was a way to change to unstabilized (or less-stabilized) video so that the viewer could get a little more of a wild ride. Or, ideally, there’d be a way for the drone to provide you with a visualization of what it was doing using the data collected by its cameras. That’s probably wishful thinking, though. The drone itself doesn’t record audio because all you’d get would be an annoying buzz, but the app does record audio, so the audio from your phone gets combined with the drone video. Don’t expect great quality, but it’s better than nothing.

Skydio 2 App
The app is very simple compared to every other drone app I’ve tried, and that’s a good thing. Here’s what it looks like:

Image: Skydio

Trackable subjects get a blue “+” sign over them, and if you tap them, the “+” turns into a spinny blue circle. Once you’ve got a subject selected, you can choose from a variety of cinematic skills that the drone will execute while following you.

You get the controls that you need and the information that you need, and nothing else. Manual flight with the on-screen buttons works adequately, and the double-tap to fly function on the phone works surprisingly well, making it easy to direct the drone to a particular spot above the ground.

The settings menus are limited but functional, allowing you to change settings for the camera and a few basic tweaks for controlling the drone. One unique setting to the Skydio 2 is the height floor—since the drone only avoids static obstacles, you can set it to maintain a height of at least 8 feet above the ground while flying autonomously to make sure that if you’re flying around other people, it won’t run into anyone who isn’t absurdly tall and therefore asking for it.

Trackable subjects get a blue “+” sign over them in the app, and if you tap them, the “+” turns into a spinny blue circle. Once you’ve got a subject selected, you can choose from a variety of cinematic skills that the drone will execute while following you, and in addition, you can select “one-shot” skills that involve the drone performing a specific maneuver before returning to the previously selected cinematic skill. For example, you can tell the drone to orbit around you, and then do a “rocket” one-shot where it’ll fly straight up above you (recording the whole time, of course), before returning to its orbiting.

After you’re done flying, you can scroll through your videos and easily clip out excerpts from them and save them to your phone for sharing. Again, it’s a fairly simple interface without a lot of options. You could call it limited, I guess, but I appreciate that it just does a few things that you care about and otherwise doesn’t clutter itself up.

The real limitation of the app is that it uses Wi-Fi to connect to the Skydio 2, which restricts the range. To fly much beyond a hundred meters or so, you’ll need to use the controller or beacon instead.

Skydio 2 Controller and Beacon

Photo: Evan Ackerman/IEEE Spectrum

While the Skydio 2 controller provides a better hands-on flight experience than with the phone, plus an extended range of up to 3.5 km, more experienced pilots may find manual control a bit frustrating, because the underlying autonomy will supersede your maneuvers when you start getting close to objects.

I was looking forward to using the controller, because with every other drone I’ve had, the precision that a physically controller provides is, I find, mandatory for a good flying experience and to get the photos and videos that you want. With Skydio 2, that’s all out the window. It’s not that the controller is useless or anything, it’s just that because the drone tracks you and avoids obstacles on its own, that level of control precision becomes largely unnecessary.

The controller itself is perfectly fine. It’s a rebranded Parrot Skycontroller3, which is the same as the one that you get with a Parrot Anafi. It’s too bad that the sticks don’t unscrew to make it a little more portable, and overall it’s functional rather than fancy, but it feels good to use and includes a sizeable antenna that makes a significant difference to the range that you get (up to 3.5 kilometers).

You definitely get a better hands-on flight experience with the controller than with the phone, so if you want to (say) zip the drone around some big open space for fun, it’s good for that. And it’s nice to be able to hand the controller to someone who’s never flown a drone before and let them take it for a spin without freaking out about them crashing it the whole time. For more experienced pilots, though, the controller is ultimately just a bit frustrating, because the underlying autonomy will supersede your control when you start getting close to objects, which (again) limits how useful the controller is relative to your phone.

I do still prefer the controller over the phone, but I’m not sure that it’s worth the extra $150, unless you plan to fly the Skydio 2 at very long distances or primarily in manual mode. And honestly, if either of those two things are your top priority, the Skydio 2 is probably not the drone for you.

Photo: Evan Ackerman/IEEE Spectrum

The Skydio 2 beacon uses GPS tracking to help the drone follow you, extending range up to 1.5 km. You can also fly the with the beacon alone, no phone necessary.

The purpose of the beacon, according to Skydio, is to give the drone a way of tracking you if it can’t see you, which can happen, albeit infrequently. My initial impression of the beacon was that it was primarily useful as a range-extending bridge between my phone and the drone. But I accidentally left my phone at home one day (oops) and had to fly the drone with only the beacon, and it was a surprisingly decent experience. The beacon allows for full manual control of a sort—you can tap different buttons to rotate, fly forward, and ascend or descend. This is sufficient for takeoff, landing, to make sure that the drone is looking at you when you engage visual tracking, and to rescue it if it gets trapped somewhere.

The rest of the beacon’s control functions are centered around a few different tracking modes, and with these, it works just about as well as your phone. You have fewer options overall, but all the basic stuff is there with just a few intuitive button clicks, including tracking range and angle. If you’re willing to deal with this relatively minor compromise, it’s nice to not have your phone available for other things rather than being monopolized by the drone.

Skydio 2 In Flight

GIF: Evan Ackerman/IEEE Spectrum

Hand takeoffs are simple and reliable.
Click here for a full resolution clip.

Starting up the Skydio 2 doesn’t require any kind of unusual calibration steps or anything like that. It prefers to be kept still, but you can start it up while holding it, it’ll just take a few seconds longer to tell you that it’s ready to go. While the drone will launch from any flat surface with significant clearance around it (it’ll tell you if it needs more room), the small footprint of the battery means that I was more comfortable hand launching it. This is not a “throw” launch; you just let the drone rest on your palm, tell it to take off, and then stay still while it gets its motors going and then gently lifts off. The lift off is so gentle that you have to be careful not to pull your hand away too soon—I did that once and the drone, being not quite ready, dropped towards the ground, but managed to recover without much drama.

GIF: Evan Ackerman/IEEE Spectrum

Hand landings always look scary, but the Skydio 2 is incredibly gentle. After trying this once, it became the only way I ever landed the drone.
Click here for a full resolution clip.

Catching the drone for landing is perhaps very slightly more dangerous, but not any more difficult. You put the drone above and in front of you facing away, tell it to land in the app or with the beacon, and then put your hand underneath it to grasp it as it slowly descends. It settles delicately and promptly turns itself off. Every drone should land this way. The battery pack provides a good place to grip, although you do have to be mindful of the forward set of props, which (since they’re the pair that are beneath the body of drone) are quite close to your fingers. You’ll certainly be mindful after you catch a blade with your fingers once. Which I did. For the purposes of this review and totally not by accident. No damage, for the record.

Photo: Evan Ackerman/IEEE Spectrum

You won’t be disappointed with the Skydio 2’s in-flight performance, unless you’re looking for a dedicated racing drone.

In normal flight, the Skydio 2 performs as well as you’d expect. It’s stable and manages light to moderate wind without any problems, although I did notice some occasional lateral drifting when the drone should have been in a stationary hover. While the controller gains are adjustable, the Skydio 2 isn’t quite as aggressive in flight as my Mavic Pro on Sport Mode, but again, if you’re looking for a high-speed drone, that’s really not what the Skydio is all about.

The Skydio 2 is substantially louder than my Anafi, although the Anafi is notably quiet for a drone. It’s not annoying to hear (not a high-pitched whine), but you can hear it from a ways away, and farther away than my Mavic Pro. I’m not sure whether that’s because of the absolute volume or the volume plus the pitch. In some ways, this is a feature, since you can hear the drone following you even if you’re not looking at it, you just need to be aware of the noise it makes when you’re flying it around people.

Obstacle Avoidance
The primary reason Skydio 2 is the drone that you want to fly is because of its autonomous subject tracking and obstacle avoidance. Skydio’s PR videos make this capability look almost too good, and since I hadn’t tried out one of their drones before, the first thing I did with it was exactly what you’d expect: attempt to fly it directly into the nearest tree.

GIF: Evan Ackerman/IEEE Spectrum

The Skydio 2 deftly slides around trees and branches. The control inputs here were simple “forward” or “turn,” all obstacle avoidance is autonomous.
Click here for a full resolution clip.

And it just won’t do it. It slows down a bit, and then slides right around one tree after another, going over and under and around branches. I pointed the drone into a forest and just held down “forward” and away it went, without any fuss, effortlessly ducking and weaving its way around. Of course, it wasn’t effortless at all—six 4K cameras were feeding data into the NVIDIA TX2 at 30 fps, and the drone was processing a million points in 3D space per second to plan the safest path while simultaneously taking into account where I wanted it to go. I spent about 10 more minutes doing my level best to crash the drone into anything at all using a flying technique probably best described as “reckless,” but the drone was utterly unfazed. It’s incredible.

What knocked my socks off was telling the drone to pass through treetops—in the clip below, I’m just telling the drone to fly straight down. Watch as it weaves its way through gaps between the branches:

GIF: Evan Ackerman/IEEE Spectrum

The result of parking the Skydio 2 above some trees and holding “down” on the controller is this impressive fully autonomous descent through the branches.
Click here for a full resolution clip.

Here’s one more example, where I sent the drone across a lake and started poking around in a tree. Sometimes the Skydio 2 isn’t sure where you want it to go, and you have to give it a little bit of a nudge in a clear direction, but that’s it.

GIF: Evan Ackerman/IEEE Spectrum

In obstacle-heavy environments, the Skydio 2 prudently slows down, but it can pick its way through almost anything that it can see.
Click here for a full resolution clip.

It’s important to keep in mind that all of the Skydio 2’s intelligence is based on vision. It uses cameras to see the world, which means that it has similar challenges as your eyes do. Specifically, Skydio warns against flying in the following conditions:

Skydio 2 can’t see certain visually challenging obstacles. Do not fly around thin branches, telephone or power lines, ropes, netting, wires, chain link fencing or other objects less than ½ inch in diameter.
Do not fly around transparent surfaces like windows or reflective surfaces like mirrors greater than 60 cm wide.
When the sun is low on the horizon, it can temporarily blind Skydio 2’s cameras depending on the angle of flight. Your drone may be cautious or jerky when flying directly toward the sun.

Basically, if you’d have trouble seeing a thing, or seeing under some specific flight conditions, then the Skydio 2 almost certainly will also. It gets even more problematic when challenging obstacles are combined with challenging flight conditions, which is what I’m pretty sure led to the only near-crash I had with the drone. Here’s a video:

GIF: Evan Ackerman/IEEE Spectrum

Flying around very thin branches and into the sun can cause problems for the Skydio 2’s obstacle avoidance.
Click here for a full resolution clip.

I had the Skydio 2 set to follow me on my bike (more about following and tracking in a bit). It was mid afternoon, but since it’s late fall here in Washington, D.C., the sun doesn’t get much higher than 30 degrees above the horizon. Late fall also means that most of the deciduous trees have lost their leaves, and so there are a bunch of skinny branches all over the place. The drone was doing a pretty good job of following me along the road at a relatively slow speed, and then it clipped the branch that you can just barely see in the video above. It recovered in an acrobatic maneuver that has been mostly video-stabilized out, and resumed tracking me before I freaked and told it to land. You can see another example here, where the drone (again) clips a branch that has the sun behind it, and this clip shows me stopping my bike before the drone runs into another branch in a similar orientation. As the video shows, it’s very hard to see the branches until it’s too late.

As far as I can tell, the drone is no worse for wear from any of this, apart from a small nick in one of the props. But, this is a good illustration of a problematic situation for the Skydio 2: flying into a low sun angle around small bare branches. Should I not have been flying the drone in this situation? It’s hard to say. These probably qualify as “thin branches,” although there was plenty of room along with middle of the road. There is an open question with the Skydio 2 as to exactly how much responsibility the user should have about when and where it’s safe to fly—for branches, how thin is too thin? How low can the sun be? What if the branches are only kinda thin and the sun is only kinda low, but it’s also a little windy? Better to be safe than sorry, of course, but there’s really no way for the user (or the drone) to know what it can’t handle until it can’t handle it.

Edge cases like these aside, the obstacle avoidance just works. Even if you’re not deliberately trying to fly into branches, it’s keeping a lookout for you all the time, which means that flying the drone goes from somewhat stressful to just pure fun. I can’t emphasize enough how amazing it is to be able to fly without worrying about running into things, and how great it feels to be able to hand the controller to someone who’s never flown a drone before and say, with complete confidence, “go ahead, fly it around!”

Skydio 2 vs. DJI Mavic

Photo: Evan Ackerman/IEEE Spectrum

Both the Skydio 2 and many models of DJI’s Mavic use visual obstacle avoidance, but the Skydio 2 is so much more advanced that you can’t really compare the two systems.

It’s important to note that there’s a huge difference between the sort of obstacle avoidance that you get with a DJI Mavic, and the sort of obstacle avoidance that you get with the Skydio 2. The objective of the Mavic’s obstacle avoidance is really there to prevent you from accidentally running into things, and in that capacity, it usually works. But there are two things to keep in mind here—first, not running into things is not the same as avoiding things, because avoiding things means planning several steps ahead, not just one step.

Second, there’s the fact that the Mavic’s obstacle detection only works most of the time. Fundamentally, I don’t trust my Mavic Pro, because sometimes the safety system doesn’t kick in for whatever reason and the drone ends up alarmingly close to something. And that’s actually fine, because with the Mavic, I expect to be piloting it. It’s for this same reason that I don’t care that my Parrot Anafi doesn’t have obstacle avoidance at all: I’m piloting it anyway, and I’m a careful pilot, so it just doesn’t matter. The Skydio 2 is totally and completely different. It’s in a class by itself, and you can’t compare what it can do to what anything else out there right now. Period.

Skydio 2 Tracking
Skydio’s big selling point on the Skydio 2 is that it’ll autonomously track you while avoiding obstacles. It does this visually, by watching where you go, predicting your future motion, and then planning its own motion to keep you in frame. The works better than you might expect, in that it’s really very good at not losing you. Obviously, the drone prioritizes not running into stuff over tracking you, which means that it may not always be where you feel like it should be. It’s probably trying to get there, but in obstacle dense environments, it can take some creative paths.

Having said that, I found it to be very consistent with keeping me in the frame, and I only managed to lose it when changing direction while fully occluded by an obstacle, or while it was executing an avoidance maneuver that was more dynamic than normal. If you deliberately try to hide from the drone it’s not that hard to do so if there are enough obstacles around, but I didn’t find the tracking to be something that I had to worry about it most cases. When tracking does fail and you’re not using the beacon, the drone will come to a hover. It won’t try and find you, but it will reacquire you if you get back into its field of view.

The Skydio 2 had no problem tracking me running through fairly dense trees:

GIF: Evan Ackerman/IEEE Spectrum

The Skydio 2 had no problem chasing me around through these trees, even while I was asking it to continually change its tracking angle.
Click here for a full resolution clip.

It also managed to keep up with me as I rode my bike along a tree-lined road:

GIF: Evan Ackerman/IEEE Spectrum

The Skydio 2 is easily fast enough to keep up with me on a bike, even while avoiding tree branches.
Click here for a full resolution clip.

It lost me when I asked it to follow very close behind me as I wove through some particularly branch-y trees, but it fails more or less gracefully by just sort of nope-ing out of situations when they start to get bad and coming to a hover somewhere safe.

GIF: Evan Ackerman/IEEE Spectrum

The Skydio 2 knows better than to put itself into situations that it can’t handle, and will bail to a safe spot if things get too complicated.
Click here for a full resolution clip.

After a few days of playing with the drone, I started to get to the point where I could set it to track me and then just forget about it while I rode my bike or whatever, as opposed to constantly turning around to make sure it was still behind me, which is what I was doing initially. It’s a level of trust that I don’t think would be possible with any other drone.

Should You Buy a Skydio 2?

Photo: Evan Ackerman/IEEE Spectrum

We think the Skydio 2 is fun and relaxing to fly, with unique autonomous intelligence that makes it worth the cost.

In case I haven’t said it often enough in this review, the Skydio 2 is an incredible piece of technology. As far as I know (as a robotics journalist, mind you), this represents the state of the art in commercial drone autonomy, and quite possibly the state of the art in drone autonomy, period. And it’s available for $999, which is expensive, but less money than a Mavic Pro 2. If you’re interested in a new drone, you should absolutely consider the Skydio 2.

There are some things to keep in mind—battery life is a solid but not stellar 20 minutes. Extra batteries are expensive at $99 each (the base kit includes just one). The controller and the beacon are also expensive, at $150 each. And while I think the Skydio 2 is definitely the drone you want to fly, it may not be the drone you want to travel with, since it’s bulky compared to other options.

But there’s no denying the fact that the experience is uniquely magical. Once you’ve flown the Skydio 2, you won’t want to fly anything else. This drone makes it possible to get pictures and videos that would be otherwise impossible, and you can do it completely on your own. You can trust the drone to do what it promises, as long as you’re mindful of some basic and common sense safety guidelines. And we’ve been told that the drone is only going to get smarter and more capable over time.

If you buy a Skydio 2, it comes with the following warranty from Skydio:

“If you’re operating your Skydio 2 within our Safe Flight guidelines, and it crashes, we’ll repair or replace it for free.”

Skydio trusts their drone to go out into a chaotic and unstructured world and dodge just about anything that comes its way. And after a week with this drone, I can see how they’re able to offer this kind of guarantee. This is the kind of autonomy that robots have been promising for years, and the Skydio 2 makes it real.

Detailed technical specifications are available on Skydio’s website, and if you have any questions, post a comment—we’ve got this drone for a little while longer, and I’d be happy to try out (nearly) anything with it.

Skydio 2 Review Video Highlights
This video is about 7 minutes of 4K, 30 fps footage directly from the Skydio 2. The only editing I did was cutting clips together, no stabilization or color correcting or anything like that. The drone will record in 4K 60 fps, so it gets smoother than this, but I, er, forgot to change the setting.

[ Skydio ] Continue reading →

Posted in Human Robots