YaCuAi

We are happy to make a big announcement!

1 May is now a special day for us. On this day our one and only automatic scrubber dryer Unit was born.

What’s so special about it? That it’s fully automatic and self-operating.

For those who haven’t been with us from the start: 2 years ago we started to develop our robot. We made a decision to think outside the box and make it from scratch. Through these years we explore the world of self-driving technologies.

And now we are ready to start the new era with Unit!

Keep up with us and maybe you’ll be the first happy owner

What in total?

Now we have a stable, unpretentious product that can move by following virtual marks and stop when an obstacle occurs, detecting the obstacle with a standard parktronic. At the same time, we spent a long time making up and writing our own smart grid. One thing that sets us apart from the big corporations is creativity :). We came up with, as all startups say, a unique neural network that learns from evolutionary algorithms that consider a variety of species (patterns). 

The main feature is that it is stable in advance in an open world conditions (Open Set), when you have to face something unpredictable, something that was not in the training. 

The second feature is the ability to remember previously learned skills we were able to overcome "catastrophic forgetting""!

The third feature: we can see how the neural network learns. On the first release, on the MNIST dataset, it made 0.25% of errors. We've friended it with our Toyota Prius, and the network is learning to measure heading angle and gas/brake levels from the video stream.

To provide the neural network with relevant video data, we developed a time-domain, a video stream filter that suppresses static, low-motion image areas, and also developed convolutional spatial filters suitable for this purpose.

Positioning system
What was done:

* pairing with a computing platform;

* iterative algorithm of multilateration;

* Kalman filters for the final computation of spatial data;

* provided resistance to disruption of the connection with the anchors, up to their total absence, depending on the duration of the disruption.

At the same time, we purchased a Robo-Machine based on the Raspberry Pi microcomputer with the minimal necessary executive peripherals, a camera, infrared and ultrasonic sensors. It is an entry-level robot, not armed with any inertial systems, and moves like a tank. 

When you do projects like this, it's important to think ahead, but you're constantly wasting time on things you hadn't even thought about. For example, it took us a long time to understand why the robot's behavior changes over time, even on simple routes in the form of circles or rounded squares.

It turned out that, due to excessive simplicity, the supply voltage to its PWM drivers is not stabilized, and the base speed "floats", depending on the battery charge. Also, in the absence of inertial sensors, it's pretty difficult to ensure proper turning angles, as well as stable straight-line motion and a lot of other things

What was done:
* developed a mathematical model;

* integrated the RTLS system;

* applied an iterative multilateration algorithm;

* implemented the Kalman adaptive filtering algorithm;

* developed an automatic control algorithm for the " free " movement, and for the route guidance;

* tested it successfully.

Refinements

Refinements were needed to calculate and estimate not only the point location (multilateration), but also to control the orientation in space. The main difficulty was that the same set of position marks (points) corresponds to a mathematically infinite number of trajectories. And exactly on time everything was ready. When shipping to Russia, we were very worried that it would be wrong, would work wrong or would not work at all. It took about 4000$ in money, but their devices exceeded our expectations in all parameters.

After working with the code, we have our own algorithms and methods of calculating spatial data, and also a positioning system, which we can cheaply enough, in a few days, place anywhere. We know how to create 3D maps of space and draw routes on them, which our platform will move through.

Iron

We were looking for hardware developers to determine the location of the platform in space. This system is called RTLS, basically it is the same gps/gnss, only autonomous, independent, has higher accuracy and can be used indoors. The system consists of fixed anchors (or beacons) with well-known coordinates and moving tags. It must provide extraction of measurement data and calculations locally on the tags, i.e. on transport platforms, where they are installed.

It took us several months to find it. Everything we found in Russia did not seem to be right for us. We met with different companies, but they all seemed like dinosaurs from the Soviet Union: a lot of patents, but when you ask about the purpose of the box, it turns out that they do not really know what they made it for.

In the end, we found developers from China, who self-organized themselves in a chat room with all the departments: commerce and development. We quickly came to the conclusion that their product was right for us, but it needed to be fine-tuned for our needs.

Actually! Already approaching the 3rd step of creation our favorite, autonomous cleaning robot. But I will write step by step, i.e. about the 2nd stage.

Most of the work went unnoticed for me and in fact was devoted to the development of the technical project (below I will give a list of works). I have mastered a little bit of experience related to industrial design.

There is a methodology by which a designer is selected.
The 1st step is to identify 3 candidates.
To them:
- they are told that there will be 3 stages, at each stage the work can be stopped if the client finds further participation unreasonable;
- the first work is done in pencil;
- everyone should make at least 10 sketches of functional design;
- the object must be convenient, understandable and fulfill the key functions;
- deadline is 24 hours, the next day the criticism and selection is formed.
In the 2nd stage, technical corrections are made, the beauty and detailing of the directions are done. Still in pencil.
At the 3rd stage, when all the functionality is tightened as it should be and there is nothing left to cut, designers begin to draw colors, fractures, inscriptions, artificial complexity.
And only after that, the 1st finalist is taken to 3D for stretching the beauty of the frame design. Since the frame is no longer changed by the customer, it is often necessary to redraw the beautiful 3D with the current inputs and it does not turn out well. Then other 3Ds are looked at. For each 3D the designer has to give out three renders.

Exactly in my case, I spent a month drawing different shapes, found the right one, added colors and gave it to the designer, who brought it into 3D and rendered it.

In general, the design of the cleaning robot is inspired from the area of streetcar technology. This means that there is an understanding that it is a moving vehicle. You can see where it's facing and you can see where it's supposed to go.
There was a lot of debate about the location of the attachments, especially the lidar, because:
- it needs the maximum viewing angle (ideally 360), so it was beaten as a "mouth";
- a stereo camera in the form of a "nose";
- LED lines, or " eyes "

Two goals for self-driving car implementation
In adopting unmanned technology, we set ourselves two goals. The major task is to make a self-driving system for long-distance deliveries. The minimum task was to automate logistics in closed territories on the basis of electric cargo trucks or passenger shuttles.
We were aware of the complexity of these tasks and how much money we would need to invest, but still we decided to go for it and with our most modest budget we managed to get pretty far. You can read how it all began here
The plan-minimum was born from the realization that it is not necessary to spend a lot of time on RnD. At the first stage it is enough to make a simple platform that will move along the virtual markers (rails) and stop when an obstacle is noticed. Later we added a neural network that explores the surrounding world and makes decisions. The problem with all the occasional platforms is that they try to put everything into them at once: neuronics, lidars, stereo cameras. This makes it extremely complicated and, in fact, ruins the project.

Self-driving cleaning robot
In my posts, I gave a summary of 2 months of looking for a contractor to do the iron and engineering work for our robot. The price tags were: over 0,2 millions $ for engineering but without prototyping, and the option we chose - around 0,1 millions $ for building a ready-to-use robot from zero point. Let me remind you that the technology of self-driving transportation already exists. It's based on RTLS (a miniature version of GPS which we can install on any terrain). The advantage is that when working in commercial areas it's dangerous to rely on lidar alone for mapping, as there are a lot of glass and reflecting surfaces. All we have to do is make a product to suit our requirements.

So, we've finished the first stage of building a self-driving cleaning robot. We analysed the competitors, made a comparison table, took apart the test subject, the Karcher brand and made conclusions about:
-size and weight
-power of motors, hoover, rotation speed of shafts
-cleaning area and the size of the bins
-battery capacity and much more

Based on this data, a sketch was made in 2 versions:

1. With rear-drive wheels on worm gear reducers.
Advantages: The tank method of travel gives a turn on the spot, better flexibility and the ability to clean corners, smaller size, tested motors, the brush is pressed by the weight of the machine and does not require additional mechanisms for their lifting.
Drawbacks: Rear wheels are always in the wet, slip chances are high. The wheels are locked, which prevents manoeuvring. There were doubts about the constant, damaging torque of the brushes, which would drift off track, requiring constant steering. A low-level programming solution was found
2. With front drive and swivelling wheel at the same time, similar to loaders.
Advantages: Drive in front of the brushes, good traction in the dry area. Disadvantages: Expensive, no experience with these wheels, no proven suppliers, have to buy from AliExpress, insufficient steering, longer base, requires a brush lifting mechanism

Decided! Option 1! Next is stage 2, design and layout of attachments. I want to try and do it myself

Hi, I'm Robert, the founder of the YaCu service. We have made a platform that solves all logistics problems. Starting from SaaS to automate your rolling stock using an algorithm to build an optimal route, to fulfillment services and delivery to marketplaces. All of these features are in one personal account.

Thanks to our strong IT competence, we spun off a small department, and we have been developing unmanned technology for more than two years already. Today I want to tell you which logistics problems this solves and what difficulties we have encountered.