Teaching robots to talk over 4G
Mar. 28, 2017 - When we were building the hardware architecture of the new warehouse in Andover, we realized we had developed a world first in radio design.
The new warehouse solution designed for our next generation of CFCs has thrown up lots of interesting technical challenges along the way. One of the first was how to communicate with over 10,000 robots concurrently.
More specifically, how do we get timely information on the locations of thousands of robots so we can efficiently control them? The answer lies in our unending striving for new and novel solutions to technical challenges.
Here is the story of how we taught robots to talk over 4G - and created a number of world firsts in the process.
Why Wi-Fi was a no-go
We started this project by studying the properties of Wi-Fi. The standard Wi-Fi network uses a distributed coordination function which is not a deterministic means of guaranteeing latency. The analogy often used is polite conversation at a dinner party: it is easy to have one single conversation at a table of two, much harder once you get to 12. Now imagine a state banquet; not everyone will have a chance to share their important information. There is provision in the standard for a point coordination function but nothing available on the open market supported this feature.
There is also the issue of scale, as most Wi-Fi access points (APs) can support a maximum of somewhere between 64 and 256 concurrent clients. Consider that we need to communicate with swarms of thousands of robots, talking to each 10 times a second and with guaranteed latency. We needed to understand what this meant for the network design as well as how to accommodate it in an already busy spectrum plan.
Through our experience of existing systems, we knew we should enable a maximum of around 80 clients per AP and reduce the cell size to a very small area. This is similar to the approaches taken in high density deployments such as sports stadiums and auditoriums. However, unlike in those areas, the clients would be constantly moving and using the Wi-Fi network as their primary activity, rather than watching the sports game or concert. There is relatively little headroom above the bots, providing a further driver for many small cells.
Roaming would be required between the Wi-Fi APs – this is notoriously unpredictable. The best results we see in the real world are around 300ms, which would impact the command and control of traffic.
Another aspect to consider is that this approach would require complex planning and installation work for each warehouse, which would be expensive and difficult to maintain.
Ultimately, we found that Wi-Fi’s non deterministic performance would hinder our ability to develop an efficient real time control algorithm for the bots. Resolving this problem was therefore critical to the success of the overall project.
Working with Cambridge Consultants
We had a few ideas for solutions, but wanted to get a fresh perspective on the problem. Cambridge Consultants with its world-class RF (radio frequency) and DSP (digital signal processing) skills were the obvious choice.
It also has the know-how and complete range of equipment onsite to manufacture small batches of the final product; in fact we learned a great deal from them about electronics manufacturing.
The partnership has proven to be very successful because we produced a fully featured prototype within two years (it usually takes three to five). In addition — and maybe quite surprisingly to some — the initial system architecture we defined at the very start of the project remains unchanged.
The system we came up with takes advantage of modern wireless communication principles but has secret ingredients that tailors it to our environment. For example, it works in license-free spectrum so we can deploy it at a moment’s notice.
In fact, it’s the first deployment anywhere in the world to use the unlicensed 4G spectrum for warehouse automation. By making the system private, we were able to improve the performance and simplify overall complexity by minimizing handshake and eliminating roaming aspects. The system guarantees a connection 10 times a second to each of the 1,000 client hosts per base station — all working within a 150-metre radius.
Like all extremely low-latency, real-time systems, we also needed to include features to provide redundancy. The devices have physical attributes such as dual network and power links and also logical tools to enable failover to hot standby units; this has enabled us to upgrade the base station firmware without impacting its operation. These tools also enable us to automatically recover from network and power outages, which has been invaluable during site commissioning tests.
Through this project, we’ve also bought Precision Time Protocol into the business; one benefit is that it enables very tightly synchronized log capture across multiple machines, which was beneficial during the development of the system.
Building and deploying this system from the drawing board in only three years was impressive in anyone’s book but we haven’t finished yet. By designing this system ourselves, we’re free to alter and add features as we see fit. Soon we’ll be analyzing how alternate MAC layers could help improve warehouse efficiency and designing a custom roaming algorithm optimized for our specific application.
This is only one of several wireless projects that are defining new applications for wireless technology – with OSP, we need to expand our capabilities in the RF domain.
Because there are many alternative applications for our scalable solution — factories, construction sites, airfields, etc. — our next task is to create a demonstration system to show off the technology to interested businesses.
Strategically, there are a number of areas we can improve to make the system work even harder, for example: miniaturization, increased processing power, and beyond. It’s not in our nature to just say job done and stop innovating, and I’m excited to see where we can take the system next.
Adam Green is principal wireless engineer at Ocado.
September 10-15, 2018
AI Manufacturing 2018
September 12-13, 2018
International Robot Safety Conference 2018
October 9-11, 2018
Advanced Design and Manufacturing Expo 2018
November 14-15, 2018