Palladyne AI CEO: ‘We’ve got a better mousetrap’ for robots
PDYN, $5.89, is extremely volatile, and also one of the top-performing stocks of 2024, rising sixteen-hundred percent.
Palladyne AI used to make robots. Now, they’ve switched to making software to make existing robots smarter. It’s a better mousetrap than the approach taken by, for example, Nvidia, contends Palladyne CEO Ben Wolff. His challenge this year is to prove the software works, and that the market will pay for it.
Onstage at the Consumer Electronics Show in Las Vegas last month, Nvidia CEO Jensen Huang, flanked by a line of humanoid robots from various manufacturers, declared that “The ChatGPT moment for general robotics is just around the corner,” meaning, a period of tremendous advances in machines that look like people and do a wide variety of tasks.
This article was originally published February 25th for TL subscribers. For more articles like this, consider subscribing.
Not everyone is so sure of that, even those who make stuff to make robots work.
“It will happen one day, but there are so many challenges, not just on the software and AI side, but on power consumption,” says Ben Wolff, founder and CEO of Palladyne AI, in a chat he and I had via Zoom last month.
“I mean, how useful is a humanoid robot if it can only operate between charges for two or three hours?”
Wolff knows firsthand the challenges. For nearly a decade, his company tried to build and sell robots of a quasi-humanoid type, an “exoskeleton” that could enhance a human worker’s activity. A year ago, it became clear that approach was way too capital-intensive.
Wolff returned from retirement to refashion the company he helped found. He has jettisoned hardware, drastically reduced headcount and operating expenses, and re-focused the company on its unique software offering.
That software, using AI techniques, represents a novel way to train today’s existing, in-service robots to be masters of not just one but many tasks, by being shown examples by a human.
The interested investor should regard Palladyne with eyes open: It is a “development-stage” company, with de minimus revenue and no sell-side research coverage. It will take time for Palladyne to build out the marquee customers it hopes to have in robotics.
It’s a stock that’s volatile, but also one of the best performers of 2024, rising sixteen-hundred percent.
For the average investor, Palladyne is one to keep an eye on as its business develops. For an ambitious investor, it may even be an intriguing gamble right now.
FIELD OF ROBOT DREAMS
Not only is Palladyne not a new story, it is one with a quixotic past spanning decades.
The original company was spun out of the University of Utah in 1983, then bought by Raytheon in 2007, and, seven years later, spun out again by Wolff and chief innovation officer, Fraser Smith.
A year and a half ago, I profiled Palladyne, then known as Sarcos Technology and Robotics, interviewing then-CEO Kiva Allgood, who has since left the company.
At the time, Sarcos seemed on the verge of selling humanoid robots with prices ranging from $127,000 to $590,000. The company, however, was losing tens of millions of dollars.
See also:
Sarcos Technology CEO: ‘We will change the industrial landscape’, April 9th, 2023.
Growth expectations soon turned out to be a mirage. Allgood’s last forecast to the Street, in May of 2023, shortly after she and I spoke, was for sales of $23 million that year. It ended up being just $6 million.
The board asked Wolff to come back full time at the end of 2023. He soon realized it was time to “put all the hardware on the shelf” and stop the bleeding.
The “Guardian XT” was one of ten humanoid robots that Sarcos was trying to commercialize. “We were spread too thin,” says Wolff of the venture into hardware. Most of that stuff has been written down and sits on the shelves in the warehouse.
The company had spread itself too thin, he observes, trying to commercialize ten different models of robot. “We were burning six or seven million [dollars] a month” at the time, he recalls. “And, we only had $40 million or so on the books, after having raised hundreds of millions” in a 2021 initial public offering.
The software, rather than the hardware, was the key, Wolff decided.
“We realized that the artificial intelligence platform that we originally conceived of, to be able to teach the exoskeleton how to become an autonomous humanoid robot, was something that we had made a lot of progress on,” Wolff tells me. “We realized it could have relevance beyond just our own internally developed hardware platforms.”
The software has been redirected away from inventing ambitious new kinds of robots to instead serving two broad classes of existing robots: factory systems already deployed, and drones used by the military.
For factory robots, the software, “Palladyne IQ,” is designed to make the machines smarter “so that they can be trained more quickly, and trained to perform a wide range of tasks instead of just a single task,” says Wolff.
A second version of the software, called “Palladyne Pilot,” is intended to shift drone operation “from being human-controlled in every aspect of their movement to being semi-autonomous, or fully autonomous,” says Wolff.
A DIFFERENT KIND OF ROBOT SOFTWARE
Nvidia’s Huang remarked at Vegas that “the critical capability is how to train these robots,” and Wolff agrees.
The Palladyne software is also meant to “teach” robots, and it has some similarities to Nvidia’s software, COSMOS, notes Wolff.
COSMOS is a simulation engine that runs a model on the computer to predict robot motion. Likewise, “We use the notion of a digital twin to, first and foremost, validate whether that package of software and hardware that we put together can actually do what we want it to do, can do the job,” Wolff tells me.
To Wolff, the software is proof the company’s struggles these many years have not been in vein. “We wouldn't be able to be accomplishing what we're doing in artificial intelligence today if it wasn't for that twenty, thirty years worth of knowledge that we developed,” he says. “If we were just ten smart folks that came out of grad school with PhDs in AI, I don't think we could be doing what we're doing today.”
Beyond that, “we have a pretty big departure” from Nvidia, he says. Huang emphasized in Vegas that COSMOS will be fed with massive amounts of AI training data gathered from actual robots, in order to produce even more massive amounts of simulated data for further training. “They’re more of a brute-force approach,” says Wolff of Nvidia.
Palladyne’s software is more parsimonious. The company has found ways to use far less data than most AI by being focused on particular tasks. Using an approach called “success-based generalized autonomy,” the company pioneered the ability to train robots “in real time.”
Palladyne runs its software not in the cloud like Nvidia does, with giant clusters of GPUs, but instead on simpler “edge” computers sold by Nvidia, called Orin. The AI models created by Palladyne to run on Orin are written in a so-called domain-specific language, which is “narrowly tailored for the task that the robot is going to perform,” he explains.
As the real hardware robot goes through the actions of a task, controlled by Orin, the robot “is able to observe what’s going on in the environment,” he says, “and then it can learn, reason and act based on what it observes.”
Human supervision alerts the robot’s software when it fails, “so then it can learn how to do it correctly.” The software is designed “so that a typical line worker in a factory can teach a robot how to perform a new task.”
“We’ve got a better mousetrap” than Nvidia’s “very grand, massive foundation-model” approach, Wolff contends. The result is software that can automate more functions with existing robots, “versus the Nvidia approach that we’re going to automate everything.”
To Wolff, the software is proof the company’s struggles these many years have not been in vein.
“We wouldn't be able to be accomplishing what we're doing in artificial intelligence today if it wasn't for that twenty, thirty years worth of knowledge that we developed,” he says. “If we were just ten smart folks that came out of grad school with PhDs in AI, I don't think we could be doing what we're doing today.”
HOW BIG IS THE MARKET?
The market for installed robots that can be upgraded by Palladyne software is a million and a half new systems that have been put into service in the last three years, Wolff estimates. There is a basic opportunity to enhance those industrial robots, “But what I get more excited about is the opportunity to make these robots more useful for new applications,” says Wolff, “where humans are still doing this dull, boring, dangerous work.”
“If you want to be able to have a robot that can do two, three, four, five, ten tasks instead of just a single task, that's where we have the market opportunity.”
The company has a commercial version of the software for industrial robots in trials, and has “a couple of contracts” with the U.S. Department of Defense. The demonstrations show the ability to train robots to take over human tasks such as resurfacing corroded aviation parts. That is a “dangerous environment” requiring a hazmat suit, “and we can actually enhance the precision” using a robot.
It’s harder to say just how big the market is for drone software. “I’ve seen all kinds of crazy estimates about how many drones there might be in the next five years for defense applications,” says Wolff. He’s reluctant to offer any concrete market projection. A commercial version of the drone software is expected “by the end of the first quarter.”
Software revenue is expected to be split between annual license fees for the industrial market, and a one-time, up-front fee for military drones. Drones don’t last long in service, so annual licenses don’t make sense, says Wolff.
IMPROVED FINANCIAL PICTURE
It’s hard to say just how far away that software revenue stream really is.
Palladyne had just under eight million dollars in revenue last year, most of it from the existing contracts for hardware and research projects that are tapering off. “Three years from now, virtually all of our revenues will come from software licensing,” Wolff tells me.
Outside of such general predictions, Wolff is reluctant to make promises. He’s told the Street about use cases and industries, as he told me, but he isn’t giving any financial forecasts at all “until we have something really material to talk about.”
“This company, unfortunately, has a history of over-promising and under-delivering,” he observes. Very true, and not just in the past couple years.
According to the investor presentation at the time of 2021 IPO, the company and its bankers gave the new entity a $1.3 billion valuation based on expected 2025 revenue of $1.23 billion and adjusted Ebitda (earnings before interest, taxes, depreciation and amortization) of perhaps $753 million this year. Obviously, nothing of the kind is in the cards.
“I’m not going to get into that trap again,” says Wolff. “No more blowing smoke.”
The current finances seem to offer runway to pursue the new software focus. From tens of millions of dollars, the burn rate has subsided to a million and a half or so per month. “There’s no reason for that to change dramatically,” he says.
“Once we really get rolling with our product, we'll probably do a little more in sales and marketing spend,” he expects, “but, generally speaking, that's kind of a reasonable run rate. You’re not going to see us do big CES shows” in Vegas.
The company did two rounds of fundraising recently, and now has forty million of cash and no debt. The hardware left over in the warehouse has mostly been written down. “We’re in good shape,” says Wolff. “The only real sin you can commit as a person running a company is to run out of capital … We’re not at risk of that.”
HOW DO YOU INVEST IN A DEVELOPMENT-STAGE COMPANY?
Obviously, investors who buy the stock are putting a lot of faith in product development while getting limited financial evidence in the meantime.
Wolff knows something about such investing challenges. He was a co-founder of Clearwire, the pioneering wireless fiber broadband company owned by Craig McCaw. That company reached a billion dollars in revenue, but no profit, before Wolff helped to sell it to Sprint and SoftBank Group.
“I think what investors want to see from a development-stage company is for us to answer two questions,” says Wolff. “Does the technology work? And, does the market care? Are customers ready, willing and able to pay for it?”
Palladyne has yet to prove either, but, “to prove that in 2025, that’s a great way of articulating what my objectives are,” he says. This year, he plans to “keep investors informed about key milestones” on that path.
Most of the stock is in the hands of insiders, including Wolff. He was already a five-percent shareholder before coming back to the job. He was granted an additional five percent in return for agreeing to work for a dollar in salary.
“It’s been a wild ride,” says Wolff of the stock’s past twelve months. “When we announced I was coming back, I think the stock was around forty-five cents.”
Just before we met, the shares had gone through a fifty-percent drop in a couple weeks. “It’s an extremely volatile stock,” Wolff concedes. (A month later, as I write this, shares sold off in Tuesday’s market downdraft by nine percent at a recent $5.89.)
In spite of volatility, “it’s a great place to be,” says Wolff of the present moment. “I was comfortably retired,” he notes. “I agreed to work for a dollar of cash because I believe in where we’re going.”
Disclosure: Tiernan Ray owns no stock in anything that he writes about, and there is no business relationship between Tiernan Ray LLC, the publisher of The Technology Letter, and any of the companies covered. See here for further details.