He reinvented how we train robots

[00:00:04]So right now this major labs approaching this industrial task wrong.

[00:00:09]They want to create like a chat GPT for robotics. You ask it something and [music] it will do the task just out of the box. But what should I do if model do something wrong? I cannot open the hood of the model and adjust things.

[00:00:23][music] So we are solving it not by pre-training and fine-tuning, but [music] we allow user to train small task-specific model themself that are adjustable. Turns out you need like >> [music] >> 1/10 of the data than the typical approach. And model become way more reliable because you can always adjust it after deployment. This is our setup what we call a micro factory.

[00:00:51]I was originally working in game dev, but I always have this feeling that it will be much more fun to make something physical. A lot of goods, a lot of packaging, big warehouse. So I started to scroll through our forums. My co-founder, he was posting about his home production.

[00:01:09]He built bicycle lights. So I sent him an email with the business idea. Hey, there is a demand for these lights. How about we just make industrial grade light and just fill this market. Yeah, so initially we assemble these devices by our own hands, then hire some people to do this and manage them. So we grew business together. Our workers were sitting behind the table and just do these repetitive tasks again and again.

[00:01:35]We saw this breakthrough with transformers with chat GPT. We have this feeling that this technology can give us great tools. So we started to experiment with things, solving basically our own problem. One of the main motivations was just have fun with new toys. So we just decided to grab our [music] stuff and move to San San What we learn is that a lot of things is not working. A lot of things you can do in better way. So, for those people who don't know, one way to control robot is like teleoperation. You have replica basically of the robot and you [music] move it and robot just repeat your movements.

[00:02:14]This is a custom teleoperation kit that we provide to our users for [music] real industrial usage. You want to be this teleoperator as light as possible. So, we have this [music] carbon rods, precise custom encoders, couple of magnets, so it stay stable and reliable.

[00:02:33]Let's say I want to train a new task when you need to insert this cable into the PCB board and [music] you need to be super precise, so this connector fit.

[00:02:42]And if you miss half a millimeter, it just break everything. Normally, you cannot do this tasks. It's too random, it's too unconstrained. So, you basically need to show it couple of different position >> [music] >> and it will generalize among among them.

[00:02:58]We have a couple of pedals just for convenience. Start with green one. So, if you move it like this, it's too unprecise. [music] So, you can push the pedal and you have scale, so you move 5 in and robot will move just 1 in >> [music] >> and in this way you can be super precise. We can just insert the connector.

[00:03:24]Yeah, so now we show it couple of examples, train it and now it will be able to generalize. So, I can just put my [music] PCB board at some point little higher and see how it performs.

[00:03:42]So, it adjust a bit and like have a smooth insertion. Let's specifically move it lower.

[00:03:50]For example, like here and let's There is a pretty common benchmark when you need to tie zip ties and it's complex task because it's a deformable object. It can end [music] up in pretty random position. So, for this task we run robot for 100 zip ties. Total success rate was 97% and it was trained from under 40 minutes of task specific data. And the next best model they require 5 to 10 hours of data for just a fine tuning stage and they have success rate of about 85%. Yeah, so all this probably pretty cool, but what's make our model really special is the ability to adjust it.

[00:04:33]So, how we do [music] this? If your robot already did a mistake, you can roll robot back in time and we let the model run the robot and we have a clutch, you can adjust and refine your robot movement in real time on the go.

[00:04:51]And then just work without me.

[00:04:57]>> That's perfect.

[00:05:00]>> So, it remember it and next time it will correct its mistakes. And for now only this approach can achieve such results. So, we have some pilots that cover wide range of tasks and for some of them you need precision. Some of them you need this nuanced movements to scoop mushrooms.

[00:05:20][music] You need to have this shaky movements to not damage your mushrooms.

[00:05:25]This customer they spent about a month trying to use depth camera, trying to program split this task in couple of areas, but still it's just not work. And then they come to us and train it their first model in like 20 minutes. Model will connect this pictures from images.

[00:05:45]As mushroom level go lower, it will scoop deeper.

[00:05:49]I actually want to show you how it was started and [music] how was our initial prototype was looking. Okay, so it was our first prototype. We decide that we need to build constrained environment, static background, static lights.

[00:06:05]Initially, we are forced to do this, but then our model became more reliable and it can just deal with slight variations.

[00:06:14]We used these 3D printed arms that [music] we designed it. Very cheap motor. We used some tricks to make these arms more reliable. For example, this motor has a lot of backlash and to compensate it, we placed it two motors with some tension that generated by opposite shifting between them. We literally push [music] these motors against each other, the same as value of backlash of these motors. Very easy to use it in our experiments with model. And then we moved on to building these arms away here. Our next step is to mass produce and currently [music] we have two products. One is the kit that you can connect to any industrial arm to use our model.

[00:06:58]And the second product is the electronic assembly cell that include these off-the-shelf robotic arms that we choose, box frame and it also includes this kit.

[00:07:11]>> Hello.