Useful Information About Warehouse Robots

Useful Information About Warehouse Robots

Hello! I'm Robert, founder of YaCu Robotics. We build the Driver software, which can automate any wheeled vehicle. We automated a Toyota Prius and test it on closed sites. We designed and built an autonomous floor scrubber for commercial premises. We work with companies that manufacture robots and integrate our Driver into them. We are also a system integrator of warehouse robots from other manufacturers. A deep understanding of autonomous technologies lets us deliver turnkey warehouse robotization.


I would like to share my knowledge of this field from a technical point of view and the experience of our work, which now spans more than 5 years. So everything below comes purely from personal experience.


A short preface. Five years ago, when autonomous technology was not yet a mainstream topic, robots and their variations were pure hype. I truly believed that the future arrives faster than it seems and that self-driving cars were only a matter of time. I shared these thoughts with a technical friend, and we decided to start exploring the subject carefully, so that when widespread robotization arrived we would at least know the material in depth. That is how we started developing the concept.


To begin with, I bought a Toyota Prius and a device that can output all the data from the CAN bus byte by byte. We needed to extract steering angles and acceleration and braking data. The Prius was chosen as the cheapest car that people had already connected to and taken over the controls of, because all its controls were electric.


Once we had learned to get data from it, we started developing a neural network and came up with a concept of focused attention, by analogy with the human eye. When we drive, we see everything with peripheral vision and switch our focus only to the events that matter. So we fitted the car with a stereo camera, connected a computer to it and started collecting training datasets. After we had learned to take full control of the car, the main question of autonomy came up: how to organise navigation. How does a car or a robot know where it is now and where it needs to go? Remember, back then we did not yet know what our final product would be.


Then as now, there is no single correct standard for navigation. But that is exactly the most interesting area for research, and it is where breakthrough technologies are hidden.


How SLAM works

How SLAM works

Robotics gave birth to a technology called SLAM — simultaneous localisation and mapping. In simple terms, a lidar scans the space around it and compares it with images (read: curves) from memory, and if they match, the robot is here! It sounds simple, but in practice there are many complications: the robot has to cross-check lidar data with wheel odometry to know how far it has travelled and whether it has arrived at the right place, and constantly make corrections.


This technology has a number of drawbacks. For example, poor orientation on large areas: too many images have to be kept in memory. Problems with a fast-changing environment: if a warehouse is filled with new pallets that were not there when the maps were first made, it is a new room for the robot and it gets lost. SLAM does not work in open areas, because the lidar has a limited range and simply cannot reach a wall to read it. And finally, decisions about how to build the route remain with the robot rather than the person.


AGV robots


It is precisely because the technology was complex and immature that AGVs became the first warehouse robots. These are robot carts that can only move along a pre-built route. The route is usually not virtual but physical: magnetic tape, QR codes on the floor or laser navigation. They are almost incapable of driving around obstacles and can only stop. This brings a number of infrastructure problems and limitations. AGVs lack flexibility and are tightly bound to infrastructure. The technology is hard to scale, because expanding the area or changing the scenario takes many person-hours. That is why only large companies with very clear and, above all, repetitive processes can afford AGVs. For example, Amazon brought the speed of delivering racks to pickers to a profitable level. But they needed it more than anyone, and they bought the company that makes such robots.


This kind of warehouse automation suits you if your processes are similar to Amazon's. A 6,000–10,000 m² (≈64,500–108,000 sq ft) area needs about 100 robots at roughly $14,500 each. Add the cost of special racks, and you will need a lot of them. You also have to consider that storage density drops, because you do not grow upwards as with conventional racking. So you need a suitable warehouse that is low but has several floors, and there are not many of those. Established manufacturers of such robots include Geek+. YaCu Robotics is a system integrator of warehouse robots from Seer, a large company with more than 20 types of robots.


About robot navigation


Understanding the problems of SLAM, we thought about how to build a robust autonomy technology. In self-driving cars localisation is handled by GNSS correction networks used by surveyors. It works like GPS, but with ground stations as well as satellites, and gives centimetre accuracy instead of the metre accuracy of GPS. But this technology only works outdoors, not inside buildings. So we decided to create an in-house analogue of GPS. We found a system of radio beacons based on RTLS, wrote tens of thousands of lines of code and got a technology where we place beacons around the perimeter of the premises (they look like ordinary Wi-Fi routers), install a tag on the robot and get not only the robot's position but also its direction of travel. This is called global navigation. Without false modesty, it is an extremely reliable system with no direct equivalent. With a pre-set map, the robot can move in any direction, to any point, over any distance. This freedom became the prerequisite for AMR robots.


AMR robots

CarGo 300 AMR robot for loads up to 300 kg

These are the next generation of robots after AGVs. They have more sensors, and the most important is the lidar. A lidar is essentially a laser rangefinder that spins in a circle to build a map and drive around obstacles. On robots the lidar is usually single-beam and costs about $110, and it gives an image in only one plane. Self-driving cars use lidars costing around $9,000 with 36 beams, which give a three-dimensional image. Combined with GNSS geodata, they make it possible to build highly accurate 3D city maps tied to coordinates, which is what mapping companies do. There are also redundant sensors, such as a stereo camera, which performs one or more functions: building a map or detecting obstacles. And the standard parking sensors all around.


This complexity is needed for SLAM, to map the whole premises and move freely within them. It makes it possible to put a robot into any warehouse to deliver anything amid people and vehicles. It will not be as fast as an AGV, but it has no movement restrictions. That is why AMR robots come in so many form factors: rack transport, a platform robot that moves components from one machine to another, or a companion that follows a picker with baskets and can leave for the packing zone at any moment while a new one arrives, so the picker never has to leave the storage area. I find this option interesting, especially as recent news says DHL is deploying such robots widely in its warehouses. An example of an AMR for moving loads is our CarGo 300 mobile robot.


Autonomous forklift

YaCu FMR autonomous forklift stacker

It uses the same autonomy technology, but a number of complications appear. First, it has to drive under pallets, so they must be perfect in terms of warping, which is hard in real-world warehouses. Second, pallets have to be lifted onto racks and placed very precisely, which requires finesse that such robots do not have yet. Third, the load on the pallet must be stacked very correctly and evenly. I do not have much faith in that scenario, but autonomous forklifts remain my favourite. They can be used to move pallets between zones, for example from receiving to storage. Obviously this reduces the time staff spend on transport. Such a robot costs close to $55,000. But we are already designing an intermediate option in which the long and unreliable calibration when picking up a pallet stays with a person. The person can lift the pallet themselves, select the destination and continue unloading the truck. Another advantage of this intermediate scenario is that the robot does not need to be integrated with a WMS.


In the next article I will talk about exactly that: what an RCS (robot control system) is and how robots receive tasks from a WMS.