Before the marketing departments of the world got ahold of the term “artificial intelligence,” it used to bring to mind some sort of robot. Boston Dynamics is one of the best-known company in the robotics space, largely thanks to its viral videos of its robots in action. But the company doesn’t talk to press much, so we jumped at the opportunity to sit down with Boston Dynamics CEO Marc Raibert at Web Summit 2019 last week.
Raibert attended the conference to talk about three robots Boston Dynamics is currently focused on, and he categorized them by time: today (Spot), tomorrow (Handle), and the future (Atlas). In our interview, he also discussed current customers, potential applications, AI, simulation, and of course those viral videos.
“Today,” for Raibert, refers to a time period that extends over the course of the next year or so. Spot is the “today” robot because it’s already shipping to early adopters. In fact, it’s only been shipping for about six weeks. Boston Dynamics wants Spot to be a platform — Raibert has many times referred to it as “the Android of robots.”
Spot, which weighs about 60 pounds, “is not an end-use application robot,” said Raibert. Users can add hardware payloads, and they can add software that interacts with Spot through its API. In fact, Raibert’s main purpose in attending Web Summit was to inspire attendees to develop hardware and software for Spot. Boston Dynamics has an arm, spectrum radio, cameras, and lidars for Spot, but other companies are developing their own sensors.
The “Spot” we’re talking about is technically the SpotMini. It was renamed when it succeeded its older, bigger brother Spot. “The legacy Spot was a research project. We’re really not doing anything with it at the moment. We just call it ‘Spot’ now; it’s the product.”
Spot can go up and down stairs by using obstacle detection cameras to see railings and steps. It also has an autonomous navigation system that lets it traverse a terrain. While Spot can be steered by a human, the computers onboard regulate the legs and balance. Spot travels at about 3 miles per hour, which is about human walking speed. It has cameras on its front, back, and sides that help it navigate, travel autonomously, and move omnidirectionally. It has different gaits (slow, walking, running, and even show-off), can turn in place, and has a “chicken head” mode. That last one means it can decouple the motion of its hand from its body, similar to how many animals can stabilize one part while the rest of the body moves.
Leasing now, selling later
Boston Dynamics plans to manufacture 1,000 Spot robots, and many have already shipped. The company expects to finish building that thousand by mid-2020. We were curious: Do businesses buy just one robot? And how much hand-holding does Boston Dynamics do?
“It’s usually not one, it’s usually a pair,” Raibert told VentureBeat. “Or, I think up to five — something like that. So we have two tracks. We have a track that we’re pretty hands-on — where we deliver the robot, we do training, and we’re kind of there to help them succeed. We want these people to succeed, because it’s going to help build the market. There is no nascent market for legged robots. We’re making it, and so this year is about helping build that market. And [in] the heavy hands-on track we might even add features if the customer needs it. Although, obviously it’s all a balance of resources and how much opportunity we think there is. And then another track is just to sell them not unfiltered, but with much less help and attention.”
In fact, Boston Dynamics is not selling Spots, yet. “Currently, we’re leasing them,” said Raibert. “I think this is the very early adopter time. There will be a time not too long from now when we start selling them outright, but right now we’re leasing them. This is the first generation of the product. I’m sure there’s going to be things that need to get fixed. We’ll see how rough customers are. So far, we’ve had very good success. We’re probably rougher with the robots in-house than they are — because we know our job is to find out what the weaknesses are and fix them.”
Raibert did not want to give a potential price tag for a Spot robot, if he were to sell one today. He did share, however, that they’re currently being leased for “a few thousand dollars a month.” And he added that “each deal is a custom deal” that pertains to how many robots a company leases, for how long, and so on.
Last week, Boston Dynamics released the latest Spot software. One improvement is more autonomous navigation. “When I say autonomous navigation, it means you can drive the robot through a site, record a map, and then tell it ‘I want you to travel this path.’ And it can do that either once or routinely. You can tell it to collect data from the cameras. It can be a payload camera or the built-in cameras.” Again, Spot is a platform. And platforms receive updates.
Spot a Spot
Industrial robots tend to be stationary. Because Spot can move and extend its arm, Boston Dynamics bets Spot can be a more effective tool as a mobile manipulation platform. The company is not interested, at least right now, in building robots that interact with humans. You’ll notice that all its robots are focused on taking care of tasks that humans will not, should not, or cannot do themselves. It’s a clear strategic decision. These aren’t robots for the home or even public spaces.
It also means Boston Dynamics can protect its IP. “The safety of the robot is dependent on keeping them reasonably away from people,” Raibert said. “Or having an agreement that people will be careful of the robot.”
Boston Dynamics is thus offering Spot only to businesses. The company currently has no interest in offering them to developers, enthusiasts, or individuals.
“We’re looking for people who are interested in developing applications where we think that there’s some growth opportunity,” Raibert explained. “So for instance, we’re working with one company that’s working in the construction space where they’ve already been selling scanning technology for progress measurement. But they have people carrying around the sensors, and we’re looking at working with them to have the robots carrying around the sensors. And there’s a lot of advantages to having a robot that can go to exactly the same places. It doesn’t require lots of humans. Some of these construction companies have thousands of simultaneous construction sites they’d like to do progress monitoring on, and they’re limited in how many people they have to do that.”
Over the next year, the most likely place that you might spot a Spot would be a warehouse, construction area, or industrial installation, Raibert told VentureBeat. And that’s not just in the United States (Boston Dynamics is based in Waltham, Massachusetts) nor Japan (Boston Dynamics is currently owned by Softbank). Spots are being shipped around the world, Raibert confirmed.
Robots can use a scanner or 360-degree camera and perform data collection at a much higher level of precision without endangering human life. At power distribution companies, for example, there are sites that, when energized, people aren’t allowed to enter, even though an inspection could be critically useful. Boston Dynamics has sent Spots into those places, on a test basis only so far.
Oil refineries are also a great use case for Spot, because of the vast amount of data collection and inspection work that needs to be done. It’s performed intermittently now, Raibert says, so doing the work more regularly could be valuable to these facilities. Spot has worked at a British Petroleum oil refinery, a National Grid power distribution plant, and a Fujita construction site, where it can potentially save money by collecting 3D scans and data routinely.
Spot also has ambitions to be an entertainment robot; Boston Dynamics has tested letting people drive Spots around for fun. Instead of controlling a virtual character in a video game, you can control a physical Spot. The company is also working with Cirque de Soleil to gauge how people and robots can work together in performances.
Police and military
There’s been plenty of talk about whether law enforcement and the military should be using AI and robots. Boston Dynamics is okay with having such customers, but there are rules.
“We’re doing a limited amount of work with police departments,” Raibert told VentureBeat. “Only on hazmat and bomb squad type activities. When they go into a situation where they don’t know what a thing is and they want to take a look? Send the robot in. They’re learning and we’re learning. It’s coming along, though.”
A couple of months ago, Raibert was on a flight from Boston to California. While he was over Chicago, he used the in-flight Wi-Fi to drive a robot around the office lab. He only had 3Hz image updates on the screen, but it was enough. While the exercise was just for fun, it proved that teleoperation at significant distances was possible. Remote operation is particularly useful for the police force, Raibert points out, because they don’t care about the autonomy part of the robots, aside for mobility.
“So we have a lot of technology for remotely operating the robot,” Raibert said. “The fact that the robot has some onboard intelligence, that can manage all its balance and its maneuvering and its obstacle avoidance, means that it’s much easier to operate because you’re not in there doing the short strokes. You don’t have to worry about which leg you’re moving. We’re developing that for manipulation too. So you can look at a screen that shows the view through the robot’s cameras, point at something, and it can either go there, or pick up the thing. So it uses the local intelligence, and then you can do that remotely.”
We also asked if Boston Dynamics has any military customers. After all, Boston Dynamics’ original BigDog was a quadruped robot designed for the U.S. military.
“We don’t currently, but we probably will,” Raibert said. “We’re trying to focus on hazard avoidance, bomb detection, and things like that. We have something in our contract where people can’t arm the robots. They’re not allowed to harm a person or intimidate them. These are not soldiers. There’s lots of things military users can use robots for: get information, deal with hazardous environments, and so on.”
“Tomorrow,” for Raibert, means in about a year and a half (i.e., 2021). Handle is the tomorrow robot because it’s still in the prototype stage. The cardboard box logistics robot uses a vision system to both locate pallets and identify the box it needs among other boxes.
“In something like 18 months, we have a robot in the pipeline which is designed to work in warehouses,” Raibert told VentureBeat. “There’s about a trillion cardboard boxes shipped around the world every year. That’s an estimate based on how much volume of shipping there is. And these things are loaded onto trucks, taken off of trucks, stacked in pallets, mixed from single SKU pallets to multi-SKU pallets, put back on trucks, shipped to retail stores, and then unpacked and sold.”
Most cardboard box work is not automated. It’s backbreaking, monotonous, unpleasant work. Boston Dynamics sees a huge opportunity to automate this task with its mobile manipulation robot. “And so we have a robot that’s only goal in life is to be able to see, pick up, and position a box.”
For those who have followed Boston Dynamics, Handle is the mobile version of Pick. “It’s funny. I don’t talk about Pick that much, but we’re shipping Pick also. It’s in one location, designed to de-palletize, and before too long we’ll have palletizing. They both use the same vision systems now. So Pick has a 3D-vision system that can see boxes, that can see pallets. We have that on Handle now so that the robot can see where the boxes are in the environment. It can localize, pick up the box it wants, go over to somewhere else, [and] tuck it in with other boxes.”
Handle is coming, but Spot is still the focus. When I asked Raibert how much of the company’s effort was focused on Spot, he said that it was about 50%. But that will quickly shift to Handle.
“I think our biggest efforts are on Spot,” Raibert said. “We’re manufacturing it outside, but we’re managing all the manufacturing. It’s the first time we’ve manufactured something at scale beyond five. So before, everything was a prototype. Now, there’s a lot of process involved in making, I said a thousand, so it’s a totally different beast. The plan is for Handle, which is a year and a half away, to use the people and the skill set that we’ve learned from Spot to manufacture it.”
Eventually, Boston Dynamics hopes to build robots with all sorts of perception systems, actuation systems, planning systems, and so on.
The future: Atlas
The “future” means “in perpetual development,” for Raibert. Atlas is the “future” robot because it’s always in the prototype stage. So, will Atlas ever ship?
“I don’t think so,” Raibert acknowledged. “To be honest, Handle was an offspring from Atlas. Because we had Atlas handling boxes and we said, ‘Well, what would be a more efficient, less costly way of doing this?’ And so, you have to look at Handle through the right lens. It’s like the humanoid [Atlas], but it’s much more simplified, [with] fewer degrees of freedom, which means it can cost less. [It can be] faster. And we’re doing a spin on Handle to have a more productized version.”
Atlas is expensive and finicky, and it requires a team of special people to keep it working. Still, its performance is higher than anything else the company has created. Atlas is Boston Dynamics’ research project: The company develops new tech in Atlas and then at some point pulls some of that tech into its own product.
“I think of it as our race car. You get all the big car companies, they have race teams. And they sort of generate a lot of excitement. They generate new technology. They’re very expensive. They’re finicky. That’s what Atlas is,” said Raibert.
Boston Dynamics uses Atlas to develop learning and optimization techniques for controlling complex behavior, but also new techniques for hardware. In fact, Boston Dynamics has put just as much work into developing specialized hardware for Atlas as it has specialized software.
The latest iteration of the humanoid robot, which weighs 165 pounds, features a hydraulic power unit that weighs about 5kg and can produce 5kw of power when the robot is operating. It also has a 3D-printed leg with hydraulic pathways, actuator boards, places for valves, spaceframes, manifolds, custom lightweight components, and so on.
Most recently, Boston Dynamics has toyed with having Atlas perform basic gymnastic routines. The goal there is to test rapid behavior development optimization techniques and machine learning, Raibert explained.
AI and simulation
Speaking of machine learning, we asked about how Boston Dynamics uses artificial intelligence in its robots. Raibert likes to divide AI into two distinct types that are most relevant to his company: Athletic AI is the equivalent of what lets us operate our bodies (stand, walk, climb, maneuver around obstacles, and so on). Scholarly AI is the equivalent of what happens when we plan in our heads (figuring out what time you need to leave your home to make it to work). Boston Dynamics is focused on the former.
“I break AI down into two parts. I call it athletic AI and scholarly AI. We’re kind of experts at athletic AI. What does that mean? It means people and animals are smart about controlling our bodies, managing the energy use of our bodies, [and] receiving the world that’s right around us and maneuvering in real time with respect to either obstacles or competitors. There’s another kind of AI where you make plans and then you execute on the plans. That’s the bulk of AI now — working at a high level where you’ve abstracted the world. You’re not interacting in real time with the physical world, you’ve abstracted the world. And we focus really on the athletic part of it. I think, though, that if you do a good job on the athletic part, which is also kind of the low-level part, you can make it easier for high-level AI to interact with you.”
In other words, it’s much easier to direct a robot to take care of a task for you if you’ve already taught the robot how to stand, walk, navigate, and so on. But Boston Dynamics isn’t only doing athletic AI.
“We’re doing some low parts of the scholarly AI. So for instance, in our warehouse robot, we have systems that know about that we have to build a pallet. We’re going to need boxes from over there, and they can schedule things. But it’s all in this very well-controlled world. It doesn’t deal with the uncertainty of [a] big world out there.”
One way of avoiding that big world when you’re building robots is to train them in simulation. And so that’s exactly what Boston Dynamics does.
“We do a lot of work in simulation. I think for a lot of the people in the AI world, simulation is good enough as the end learning device. For the kind of physical things we do, it’s always a combination of the physical experimentation on real robots and simulation. The two go hand in hand.”
The physical world is still much better for Boston Dynamics to do its work in, but simulation has cost-saving advantages.
“If you have a developer who’s only worked in simulation, they’re only going to go so far in doing our kind of work. If you’ve had someone who’s done work on physical robots, they’re much more effective in simulation because they know what the constraints are, what parts of the model are good, and what parts of the model are bad, and they can steer around those things. Simulation is faster and less expensive. Sometimes you can have less experienced people make progress in simulation before they’ve gotten all the skills,” said Raibert.
On the face of it, it would appear Boston Dynamics has an experienced PR department, but there isn’t really one at the company yet, Raibert told VentureBeat. Somehow, though, Boston Dynamics videos seem to always go viral. The videos we linked to in this article alone have over 65 million views combined. We couldn’t help but ask Raibert if this was part of a larger strategy.
“It’s been interesting how we’ve been able to build a brand just through YouTube,” Raibert said. “It’s been great for the company. We don’t have to introduce ourselves from scratch, [and we get that] at very low cost. And it’s fun. We’re really robot enthusiasts more than anything else. We’re not venture-backed businesses that are only about making money. We’re really in this because [of] the challenge of understanding the key ingredients in people and animals that let them move in the world and be so effective in the world. That’s a huge challenge. Building machines that embrace that challenge is really what our long-term goal is. And commercialization is one of the things along the way. That’s got its own excitement and rewards. It’s really fun to have it out in other people’s hands, not just in our hands, which provides, a validity test and a robustness test on the work we do that that we’re excited about too.”
In short, “when you do this kind of work, you want to show people what you’re doing, and I think that’s really the motivation,” he said.
There’s been some criticism that Boston Dynamics videos don’t show that someone is controlling the robot. The videos don’t have any technical detail. They’re clearly meant to go viral. And there are tradeoffs.
“In the early days, we showed robots doing a bunch of tasks out in the world. And some people thought that they were fully autonomous, when they weren’t. There was a driver, although there was a lot that was autonomous. We were just trying to show that a legged thing that we created could climb up a muddy hill, and that that was the message. It had never occurred to us that someone wanted us to describe all the caveats. Like, it only ran for however long it ran. Or that there was a driver and all that stuff. If you look recently, we’ve been putting a few more words in that kind of qualify how autonomous it is. But we don’t want to devolve into a technical paper. Frequently people will post robot videos that that have a narration. No one watches them. Look at how many hits they get. No one watches them because they’re drowning in technical detail that most people don’t care about.”
We asked Raibert whether there was anything else he wanted our readers to know. He said that while “a lot of the business world are opportunity-first people,” his employees are not. “I think a lot of us are lifers. We’re robot-first people.”