The Evolution of Stretch

Oh, At Boston Dynamics, we really focus on doing closed loop control for robots. What I mean by that is it's always learning from the environment and reacting to it. Stretch is our first robot that's really purpose built for the warehouse. Moves cases all day long. It doesn't get tired, and it doesn't get injured. The Stretch's morphology is a mobile base with a capable drive system to move in any direction, which houses the compute and the battery power to control the whole robot. On top of that, we call the turntable, which carries the robot's arm, a perception mast, and the vacuum system. The whole was designed together, it's not an arm bolted onto a base. It's really a single robot. The other thing that makes stretch special is that it's able to understand the environment and make its own decisions about what to do next. The robot's really aware of its state and the state of the world around it, and reacting, even as things unfolds away from the ideal. So some people will design robot trajectories that are preplanned or feed forward, and, you know, assume that everything go perfectly and you'll stay right on that trajectory. But of course, the real world doesn't usually unfold that way. Even though stretch looks very different than Atlas, a lot of the underlying technology shared. We're able to use the same technological building blocks and really create very different types of robots, whether it's a humanoid or more purpose built robot like stretch. We were doing some internal research on manipulation with Atlas, and we happened to use box for this demonstration. When we released this video, we got a lot of incoming calls that Alice would be perfect for a warehouse and we wanna buy one. We knew that it wasn't really the right fit for a product at the time, but we realized that it was an indication of a market need. We thought, well, maybe we can build a road that's a little more simple and a little more purpose built for the warehouse. And that's how we got to handle. The initial idea was to combine the best of wheels and legs. Warehouses have very large open floor space that are very flat. We wanted to understand what are the capabilities and limits of a system that has that combination of wheels and legs. So we put it through an obstacle course, classic boss aerodynamic style. Early versions of handle leveraged copies of Atlas's arms on the torso to pick up and move objects, but they weren't really designed for the kinds of ranged emotion and applications that handle was really intended to do. Once we started to think more about the workspace needed to do palette manipulation, box moving, and truck unloading, we reimagined the, an tire design into something that really suits the worst base of the applications in the warehouse. So handle has, let's say, three main components. There's the wheels that uses for mobility and for balance. There's the tail. The tail interestingly is where the battery goes, and that's by design. So as this handle is moving across or reaching across a palette to grab boxes. The tail acts as a counterbalance, so it doesn't fall forward. The third element of handle is an arm that is designed specifically for warehouses And on the end of that arm is a vacuum suction gripper, and that enables it to quickly pick up boxes. And we even took handle out into some warehouses with some early adopt partners and tried to do some real warehouse tasks. The problem was that Handle has to spend a lot of time maneuvering to really do its job. Imagine trying to, do certain things in a wireless environment when you're running roller blades. And it just slows down the speed of operation when you're trying to, you know, move the locks point into point b. It's hard to continually have to turn, go turn. We could only look where the robot body was pointed. And combining that with the constraints of balancing meant that there was a lot of extra motion of the whole body to move a box from one point to another. And when you're unloading a truck, that's a kind of repeated motion that needs to be as quick as possible. When we realized that handle was a little too slow to do the job, we quickly pivoted it to stretch. The prototype of stretch consists of three main pieces. There's the base, which is about the size of a pallet. There's the arm, which does the manipulation, and there's something we call the perception mask. So that's a tall mask next to the arm that has stretches vision system. The mast is individually, actuated. So it can look around independent of what the arm is doing. And it can see what to do next. The sensors on the perception of mask, we can point, and anything we need to see, whether it's boxes, pallets, or conveyors. And we can opportunistically take pictures whenever we need to. Those pictures are then fed into a custom built machine learning algorithm that's able to identify worthy objects of interest like boxes are in the image. The arm is a seven degree of freedom arm. So there's a little more flexibility than a traditional industrial robot arm. In order to fit into tighter spaces and have more solutions to move through the space. And then at the end, we have a vacuum gripper that allows us to attach to the box and maintain a good grasp. A mobile base was really maneuverable. If you drive in any direction, you need to see forward, backward, left, right, you could turn in place, you do combinations of those motions, so that maneuverability was really a big advantage, especially in tight spaces like the back of a truck or, you know, narrow idles in a warehouse. So once the prototype of Stretch started looking promising, we started designing the product version of Stretch. We revisited every component to make sure it was gonna be reliable, cost effective, safe, and have the performance needed for the market. What resulted is a brand new machine. One big difference is the sound. So we have a new vacuum pump It's much more quiet. The robot's a lot easier to work around. Another big difference with the product version of stretch is that it has a brand new safety system on board. So that means it has an onboard dedicated safety computer. It has dedicated sensors that are redundant enable us to be safety compliant. And it has safety rated lidar the base of the robot, but it enables it to have a really high reliable way of detecting people getting too close to the robot. The same folks that designed the Atlas robot came on board, and designed a whole new gripper that we called the smart gripper. That gripper is full of sensing and pneumatic control, and it's able to handle not just perfect boxes, but boxes that may have some damage to them. It enables us to pick up more than one box at a time. Stretch is continually learning and improvement. We're constantly analyzing data that we're getting in for moments in the field. Machine learning or especially deep learning is inherently very data hungry. So the more data you can collect, the more images you can capture, the more experience the stretching and leverage, the better off it's gonna Stretch is out in the world, moving boxes every day at multiple sites. We're watching the numbers climb. We're getting rave reviews of its ability to get the job done in the worst conditions. We hear stories of trailers pulling up full of thousands and thousands of fifty pound boxes. You know, this could be in the summertime where The temperatures are over over a hundred degrees inside this metal box. And those are grueling conditions for the the people who are asked to unload them. They're very happy to able to push a button and have stretch go in and do that job while they work on other higher value operations within the warehouse. I got into engineering robotics machine learning because I wanted to work on real problems and, you know, try to do some good in the world. It's been really cool to see the the robots evolve is to become better fit for these applications we're targeting. For a company that has made its history with exciting robots, walking around and doing back flips. You get a very different sense when you see Stretch operating. It's really interesting a few minutes, and then you realize it's boring. This thing just keeps going all day long and doing real work. It's really cool to know that things we're creating are out in the world. They're helping people. And as we scale up, there's probably a Stretch running at all times of the day.