Boston Dynamics puts its robots to work

At first glance, Boston Dynamics is a strange fit for a show like ProMat. For decades, the firm has presented a flashy image to the world — a company well known for robotic highlight reels, from the snow-traversing Big Dog to parkour-performing Atlas. But a recent approach has found it facing one of its biggest challenges to date: commercialization.

As anyone in hardware (robotics especially) will happily tell you, bringing products to market presents an entirely new set of challenges, namely: repeatability, robustness and reliability. Making a robot work for long enough to shoot a viral video is one thing. Having that robot perform tens of thousands of tasks around the clock is a different thing entirely.

Stretch is Boston Dynamics’ second commercial robot. But its path has been very different than its predecessor. A straight line can be drawn from Boston Dynamics’ DARPA-funded Big Dog to Spot, and the quadruped’s platform approach to the market means giving customers a lot of control over how the product is ultimately deployed.

Stretch, on the other hand, is purpose-built. The project shares DNA with Spot and the humanoid robot Atlas, but the system was designed with a task in mind: Namely, moving boxes around a warehouse. Its direct ancestor is Handle, Boston Dynamics’ first attempt to perform the task. As Boston Dynamics’ head of warehouse robotics Kevin Blankespoor explains below, the form factor simply didn’t fit the setting.

We sat down at ProMat this week to discuss the company’s journey into the warehouse and the road that lies ahead.

TechCrunch: Boston Dynamics didn’t have a focus on commercializing for a long time. It really happened under SoftBank.

Kevin Blankespoor: Google a little bit, but SoftBank was the big push, for sure.

Atlas’ ability to move around boxes was more of an edge case. There was interest from potential clients, but there was never a plan to commercialize.

Yeah. [Atlas] is a general-purpose robot, so if you want to go explore an application, it can do the vast majority of them, just like a person can. It may not be the right optimization when you’re said and done, but it’s how you understand the application and you can peel off and do a derivation of Atlas, basically, and go make that into a product. With Atlas and case handling, we found that to be an attractive match, and I peeled off a big part of the Atlas team to do Handle, the two-wheeled robot. That was a foray into combining wheels and legs, which we always wanted to do, but also it was a little more simple and could handle cases in a warehouse.

We actually did a few experiments with customers there and then we went to Stretch. But it all really started with Atlas. A lot of the hardware tidbits from Atlas and Spot are on Stretch. It’s totally different, but under the hood, there are similarities.

What’s an example?

If you look at the hip actuators on Spot, they’re basically the same as the wrist actuators on Stretch, but Stretch is a lot bigger. The vision system across Atlas, Stretch and Spot all use a lot of shared software. We really don’t have to start from scratch when we bite off a new product. We have a big war chest of technological building blocks.

Was Kinema a part of that?

On the box detection side, absolutely. Kinema has the best box detection machine learning pipeline out there. We’ve expanded it a lot on Stretch, but that’s absolutely what it’s running now.

Discussing truck unloading has been eye-opening. Not only is it extremely physically taxing on workers, those trailers are also exposed to the outdoors, making it extremely hot or cold inside.

It can be dangerously hot. Sometimes I’m shocked to see that they let people go in when it’s 110-degrees in there. And it gets super cold in the winter.

Is hot or cold harder?

Cold is generally a challenge. We did bigger temperature extremes in our DARPA days. When you go out in the snow, and nobody’s laptop will boot up, but your robot still does. But once you’re up and running, robots generate their own heat. That’s why there’s cooling vents all over it. The computers heat up, the actuators heat up. That works against you in hot environments.

Everyone is talking about humanoids right now. After so many years and so much money in R&D, you had a lot of potential customers expressing interest. Did you explore the humanoid form factor for what ultimately became Stretch?

We didn’t explore the route because we thought it would be somewhat more expensive and complex than it needs to be. Keep in mind, that was about seven years ago. It’s amazing that we’re able to sit here and talk about maybe four or five different companies that are going after real products with humanoid robots. That’s fantastic.

Even Agility has something humanlike.

They’re definitely in that conversation. I’m just excited that if you had asked me that five or 10 years ago, I’d so, ‘oh, that’s way too expensive. That’s way too complicated.’ And now you’re finding multiple companies are looking at this and saying, ‘we think there can be a real business here.’

But for you, the move was focusing on the best robot for a single, specific task. That’s how we ended up with Stretch. It is a multipurpose robot in that the hardware is capable of doing lots of different jobs in warehouses. We’re starting off unloading trucks as our first application. That’s our first focus, but we are also in time going to be able to do things like build pallets and load trucks. In the future, a day in the life of Stretch might be spending the morning on the in-bound side and unloading boxes from containers. And spend the afternoon in the warehouse building pallets and it might move in the evenings to the outbound side loading boxes back into trucks. It really is going to be a single robot that you can do for multiple jobs.

It sounds like that’s far off.

We’re rolling out new applications every couple of years. We’re focused on getting it out of the gate.

Why didn’t Handle work out? The counterweight system was very cool, but was it ultimately more robot than was needed for the situation?

We actually found that Handle wasn’t fast enough. We took Handle out for a couple of different trial runs in warehouses. The first application was pallet loading, which looked pretty good. The second was truck unloading. What we found with Handle in the back of a truck is it had to do a fair amount of maneuvering in the space to move a single box […] It would take about 25 seconds per box. People can move a box in maybe five seconds. We knew we needed to hit that kind of number for a return on investment. Stretch can do that. It uses the arm more completely and only moves the base when it needs to.

We were talking about Stretch moving from Point A to Point B in a warehouse. But when it’s working on a shelving system, it’s able to move along the shelves [for picking]?

Correct.

Interoperability is the big topic of conversation this week. What is Boston Dynamics’ approach to this? If you want close to a fully autonomous robot at this point, you’re going to have to work with a lot of different robots from a lot of different companies.

Absolutely. It’s a hot topic. Both heterogenous fleet management. We’re seeing a lot of different companies take that on. We have prototyped our own fleet manager to understand what the hard parts are. Spot has Scout, which is its own web client for teleoperation, industrial inspection and autonomy. So we have a version of a fleet manager now. I think we have yet to see how that’s going to shake out.

Developing your own fleet manager is a resource intensive way to understand the problem.

Yeah. I mean, part of it was that, to do the things we wanted to do, there wasn’t a fleet manager you could buy. One of the things we want to do as a future application is case pick. That’s where the robot is in the aisle and it’s building up pallets of different boxes. It’s a really big application. We’ve played with different ways to do that. One is with autonomous forklifts. It moves the pallet through the facility back and forth down the aisles and Stretch rendezvous at each point location and the pallet wrap. We’ve prototyped that. We’ve done it with a couple of different partners. We’ve done it with Stretch on its own, where it’s dragging its own pallet, just to get a feel for where the hard parts are.

But you’ve decided not to go further with that?

No, we will. We basically decided to focus on application number one right now. We want to get it to the performance, reliability and robustness that customers are happy with. That’s all hands on deck. But we’ve still got a lot of the software and knowledge from when we did the picking thing I described, and that will pick up in the not too distant future.

So we can expect to see fleet management software from Boston Dynamics?

Maybe. To be determined. We’ll definitely have that for our robots. Will we have a heterogeneous manager that will work with other robots? I don’t know.

You survey the landscape, and if someone is already doing it well, you focus your attention elsewhere.

Exactly. We don’t want to do that as a primary business. If we can buy something that really works, that would be really interesting.

You mentioned the mobile picking robot. I think that’s a magic bullet. If you’ve cracked that, you’ve cracked warehouses.

Yeah. And even more generally, the thing I’m excited about is mobile manipulation. If you move around these types of shows, you’ll see a lot of mobility. There are a lot of autonomous mobile robots and forklifts. You’ll see a lot of people doing just manipulation — everything from cobots to industrial arms. But there really is nothing out there at scale that does mobility and manipulation together.

I spoke with the CEOs of Locus and 6 River Systems. They’re really just operating in that mobility segment, and having a lot of success. But successfully mounting an arm to one of those robots in a way that works is going to be huge for this industry.

I don’t know that they have the motivation to tackle something even more complicated. They have so much market penetration to go with the existing product. But for us, it’s core to what we do. Stretch is a mobile manipulation robot. It’s in the warehouse now. I think we’re going to grow in the warehouse for a long time. It’s going to go beyond the warehouse in time. Spot with an arm is a very capable mobile manipulation robot. Atlas as well, we’re starting to get more into manipulation. For Boston Dynamics, that really is the evolution, to go into mobile manipulation.

What is the upper weight threshold for Stretch’s picks?

Fifty pounds.

What are the constraints, in terms of weight?

A couple. Grasping is one. Suction is amazing. People are always amazed by how much it can pick up. But, when you start getting into imperfections like a damaged box — those top boxes, you can’t fit the gripper above it. You have to get the front face. That’s where some of the limitations come in.

If we’re talking about a mobile system that can pick, in order to compete with companies like Zebra, Locus and 6 River Systems, you’re going to have to scale down. Does that make sense for Stretch?

Yeah, absolutely. I think there are versions of something that looks like Stretch, or even a combination of Stretch and Atlas, where it’s a smaller footprint that works in a tighter space. We’re excited about all of these things.

Repeatability has to be one of the biggest issues entering the commercial space.

Yeah. Making a cool demo video is great, and I love it. I’ve done a lot of them. The difference between that and a product that works for a customer all day, every day, has sufficient hardware reliability and software robustness so it can handle all of the surprises you get — that’s an order of magnitude more work. That’s what we’ve done with Stretch. Stretch is never going to be a highlight reel, but Stretch works.