After 28 years of research and development, Boston Dynamics entered the commercial robot market last year with the launch of Spot. This is a ~$75,000 robotic dog that can march around facilities for remote inspections and, with an extra arm attachment, can even open doors and do remote manipulation.
Today, Boston Dynamics’ quest for commercialization continues with the announcement of a second commercial robot, “Stretch,” a box-moving bot designed to meet the demands of warehouses and distribution centers. The robot is designed to “go where the work is” in a warehouse, unloading trucks, de-palleting shipments, and eventually building orders. For now, we’re seeing a prototype, but Boston Dynamics hopes companies will start buying Stretch when it hits commercial deployment in 2022.
As Boston Dynamics VP of Product Engineering Kevin Blankespoor told us shortly before the launch of Stretch, the company is going to where the customers are. “When we released our first Atlas “Next-gen” video,” Blankespoor said, “there was a part of that video that showed Atlas moving boxes, and we got a big reaction from people in the warehouse space. They wanted Atlas to come work at their warehouse.” Atlas is the company’s do-everything humanoid research robot and is probably far too expensive to be a commercial product.
Blankespoor continues, “We thought, ‘Well, Atlas is probably a bit complicated to actually go work in a warehouse, but we could design a robot that’s much more simple that has the same attributes.'” The result was not just a new robot, but a whole warehouse-focused division inside of Boston Dynamics, headed up by Blankespoor.
Getting a handle on Stretch
With a clear demand for warehouse robots, Boston Dynamics started experimenting, first with its “Handle” robot. Handle started off life based around a new “wheel-legged” mobility platform—that’s two legs, but with wheels at the bottom instead of feet, enabling all sorts of nimble movement. The first version of Handle was a humanoid-ish robot that could do all kinds of crazy tricks thanks to its wheel legs, like banking around corners, rolling down steps, and leaping onto a table. The video shows Handle lifting a box with its two humanoid arms, but the robot wasn’t meant for warehouse work yet.
For version 2, Handle was “re-imagined” from what seemed like a stunt robot into a warehouse robot, and instead of two arms, it used a big vacuum gripper to lift boxes. Handle still used the wheel-leg concept but now with a big counterweight on the back, and along with the long “neck” that supported the vacuum gripper on the front, it looked more like a bird. Blankespoor says Handle could cover a lot of Atlas-in-a-warehouse use cases but with one-third as many joints.
With Handle, Boston Dynamics got as far as doing experiments with customers. The warehouse work shown in the Handle version 2 video showed the robot loading and unloading a pallet in a big, open space, and Handle was good at that. The problem with Handle is that sometimes warehouse work needs to happen in a confined space, like unloading a truck, and it was there that the robot had problems.
“It became clear that for Handle, maneuvering in a tight space was tricky,” Blankespoor explained. “We could do the job and get all the boxes unloaded, but it took too long, basically. Every time Handle picked up a box, it had to roll back to the middle of the truck so it could rotate without collisions, roll forward, and place the box.” When lifting boxes, Handle kept its balance with a big, swinging counterweight in the back, and it sounds like the robot’s big bird butt was constantly getting in the way. “We knew that there was this other option out there,” Blankespoor said. “We were considering something like Stretch for years.”
Stretch is the first Boston Dynamics robot that’s “fully purpose-built” for the warehouse, and you can see that a lot of the nimble bird design has been thrown out in favor of a big, hulking industrial robot. We’ll start with the base: the robot is simply mounted on a big box now, so it’s stable by default and doesn’t have to actively balance anymore. The robot weighs 2,650 lbs (1,200 kg) now, so there’s no need for a big, swinging counterweight when lifting—it’s not going to tip over. The arm can spin around on top of the base, so it can unload boxes from a truck to a conveyor belt without needing to move and bump into something. The result is that Stretch can unload a truck about five times faster than Handle. Stretch can move up to 800 boxes an hour.
Most warehouses are designed around the 48×40-inch dimensions of a pallet, so the base of Stretch just happens to have a 48×40-inch footprint, and it can fit anywhere a pallet fits. Wheels in each corner of the box, all with independent steering, let Stretch move in any direction, including side to side or rotating in place. The giant base also means there is a lot of room for the battery, enough to power Stretch through an eight-hour work shift, or up to 16 hours with “the extended range option.”
Going to where the work is
There are lots of stationary robot arms that can move boxes around, but fixed-infrastructure arms need costly installation, and you need one for every location where a box needs to be moved. Being mobile means Stretch can do the work of multiple stationary arms as the needs of the warehouse dictate, without the need to redesign or install anything. Blankespoor imagines a typical day in the warehouse for Stretch: “Stretch might spend the morning on the inbound side of the warehouse, unloading boxes from trucks. It might spend the afternoon in the aisles of the warehouse, building up pallets—those will go off to retailers or e-commerce centers. And it might spend the evening loading boxes back into trucks.”
Stationary arms can be as beefy as they need to be, but being mobile means Stretch needs to watch its weight. Boston Dynamics’ custom arm design is one-fourth the weight of an industrial arm, while still being able to out-lift its predecessor, with a 50-pound max payload (23 kg) versus the 33-pound (15 kg) capacity of Handle. The arm needed to be designed so it could reach across pallets and boxes all the way at the top of the truck, where there won’t be much clearance. The robot actually grabs the top row of boxes from the side, since it won’t be able to fit between the box and the roof.
The final major component of Stretch is the perception mast, a big tower that sits on the same rotating base as the arm and houses most of the robot’s sensors, so it’s never in the way of the arm. The mast houses both 2D and depth sensors, giving Stretch a high-up view of its surroundings. For vision, the robot uses Boston Dynamics’ “Pick” software, a collection of machine-learning-powered algorithms for detecting and moving boxes, which arrived at the company via an acquisition of Kinema Systems.
The base of Stretch actually has a modular interface where you can attach various accessories. For truck unloading, you can attach a conveyor belt to Stretch, so the robot can bring the conveyor belt with it as it moves deeper into the truck. This means it only ever has to just pick up a box, spin around, and drop it for faster unloading. There’s also a pallet cart attachment, so the robot can haul a pallet around as it builds orders. Additional sensors can be attached to the base, too, either for situational awareness like extra cameras or lidar, or a barcode reader for input.
Boston Dynamics hopes to sell Stretch in 2022, but it still has some iteration to do before then. What we’re seeing right now is the Stretch prototype, while the “product” version of Stretch will be sold sometime next year. As Blankespoor explains, “The Stretch product will look a lot like this, but it’s really been totally redesigned from the ground up. Every component’s been reworked for manufacturability for cost reduction, reliability, and higher performance. So the Stretch product, we will start building the first units of that this summer, and then it’ll go on sale next year. We’ll start rolling out applications that the product can do, incrementally. The first one we’ll do is truck unload, and then a little bit later we’ll start doing pallet building.” Blankespoor says the final product will get a few more sensors, like a lidar on the face of the robot.
“The other thing with the prototype is that our whole software team gets a jump start on developing control systems, vision and autonomy, and testing it on real hardware,” Blankespoor tells us, “so that when you get the product, the systems are a lot more mature.”