Cloud on Edge, robot in the future.
Notice Some information contained in this article has been intentionally obfuscated. This is to protect organization and company secrets according to internal regulations.
“I can see clearly now.”
Introduction
Cloud computing and edge computing are hot topics, each year huge investments get allocated in these areas. But few recognize the power of them when combining with robotics. There exists attempts to fuse these three aspects together but most of them end up with failures. Even with those “not failed” trails, almost none gets deployed in real world. Do not interpret wrong, it is not roboticists who do not want this fusion to happen. It was the infrastructure and regulations which forbids further advancements in the category. It remains as one of the sweetest dreams for people to explore and realize.
In this blog, several attempts would be presented to illustrate typical scenarios and possible enhancements can be made to nowadays robotic systems. Important to note some or all of the attempts may just be hypothetical assumptions or simulated use-cases. The experiments conducted in real worlds followed all regulations and laws.
Car
Autonomous driving is thriving today, what more thriving is low cost full capability autonomous driving solution. In fact, most manufactures and tier 1 OEMs are all doing something to reach higher level of AV with lower cost. One interesting idea emerged in the process is: “Why can’t we just cut the chip and have the car always connected. We already have cellular.”. Yelp, one typical case where customer knows nothing about how terribly terrible the idea is to the people who actually has to do the implementation. However, as money dumps into this research, current framework and typical implementation do exhibits some exciting behavior. From a framework point of view, such implementation requires little to be modified. Taking ROS for example, the underlying structure made sure as long as the basic RMW functionalities were implemented the system would work. The framework and upper layer algorithms get isolated from the communication abstraction. And in fact there is really no magic at all. Think about what people usually do, they put a laptop together with a robot and let the laptop do complex computations such as perception and localization. If one uses wireless connection then it is some sort of cloud already. And by setting a robot with a wireless connected laptop the simplest edge-cloud model can be obtained. And such setup is extremely common.
Of course the actual demo and experiments were much more complicated and involved a ton of prerequisites. For example, the latency and tolerance for the entire system was measured. From this step the maximum latency possible was calculated. Then the actual inference and other various algorithms were tested in the cloud with many setups while the one fitted the needs the best was chosen. Then a simulated full course test drive was conducted, and went perfectly well. So perfect which some people even think that now the company has a solution for “chip-less” AV. On the other hand, another group know the golden rule which “in theory, everything works in theory”. So a real test drive was conducted. And this is when everything immediately fell apart. The system would randomly fail to meet cycle requirements and just impossible to even drive itself onto the road. The reason on the other hand is extremely simple as many would guess, latency spikes caused by networks. No matter how good the infrastructure can be, there just exists no way to rule this out in a fashion that is acceptable for the project. But that does not mean there is really no way to sort it out. Imagine a case when one was doing only gaming and experienced unstable connections. With some games which designed smart enough, local clients would take over some computation and wait for a verification from the sever side for later updates. Further information and implementation is up to the reader’s imagination, but one thing can be made sure is such enhancement is doable in some cases to compensate for the absence of cloud side.
Inspector
Robot inspectors sound like something straight out from the BladeRunner or Matrix which humans are replaceable organic cell batteries(aren’t we not right now?). But in fact the presence of robotic inspectors nowadays is to replace jobs require nothing but patience and concentration. For example, imagine in a huge electric sub-station, people has to check the temperature of wires and instruments each hour every hour no matter what goes on. And the only thing needed was to point a thermal gun to the surface of objects with interests and record a reading. One may need around 100 people just to do such tasks every day in the middle of nowhere. Few people would like to take such jobs without a salary one can not resist. Robot inspectors would be a nature choice for such tasks. But how and where does cloud and edge blends in such scenarios, the answer would be environment and cost. In such stations, it is not uncommon to require solution vendors to do transparent operations. Which means each and every bit of data generated never goes out of the control of the station. This may due to the concerns for protecting infrastructures from terrorism attacks. Put that aside, such stations already has a huge number of edge computing nodes, the cameras, sensors, and various other equipments which go out of the imagination of people who have never stepped into them. How to properly utilize them and integrate robots into such existing framework poses an interesting challenge.
Many existing solutions would just be to use the standard routine. Robots operate in such environments as they are in any other situations but with its remote side deployed into the custom cloud. The different edges from robots and existing sensors would work independently. However, an opportunity was given to the researchers to explore other possibilities. This is essentially to experiment what can be done in a “context rich” environment. And the outcome could be pretty amazing. Imagine sensors can be integrated into the framework with a plug-and-play style. For example the framework has the capability to accept extra input from other sensors that are not pre-thought. And robots could be benefited from this information all thanks to fact that the entire back-end is more or less on the same “cloud”. And what even more surprising is ROS requires very little modification to be fitted into such scenarios. It is true that one may need some work to make sure each and every driver for the external sensor works but after that there is really no major obstacle. However, despite the high potential for external edge nodes’ fusion with robots. Truth is the actual immediate return is rather low. Not saying it is not beneficial, but it really does not matter so much for a robot to read navigate at a faster speed or point to the target more accurately. If the stations do not have those existing infrastructures, it would be very unlikely or just simply impossible for users to install the extra edge computing nodes. Fortunately, as more and more IoTs are added to the industries, people could soon use the extra benefits with no extra cost. Better is the fact that as more diversified edges deployed, there could be strong chemical reactions that may ultimately make the cost worse.
Swarm
Swarm robots are rather complicated systems in nature. And in fact this is where the cloud’s edge really gets blurred (pun intended). In many papers and other experiments one could observe that most such swarms consists only small and low cost agents yet exhibits powerful behaviors which sometimes can never be done by using any other methods. This is the perfect example of robots’ fusion of cloud and edge, each agent is both an edge computing node and combined they form a cloud which contains great information processing power. Even better is the fact the computation power maybe added by demand, for example, a “real” cloud server hosting some powerful algorithms that would optimize the behavior of the swarm as a whole. This algorithm can be preferably online but may be nearline or even offline as it serves as a supplement only or to tackle hard situations on demand. Here the algorithms are the more interesting parts, normal or classical solutions would have a hard time finding a way to find in such systems. But for some algorithms, particularly reservoir computing like systems, things can go very smoothly. The agility and ability to adapt to vastly different and fast changing resources grant this particular family great advantage in such cases. There exists many examples, and do look for the keyword of “massive parallelism”. Under this category one could experience the true power and meaning of what should be an AI. The glimpse of future they offer motivated countless people to devote their talents.
Despite the existing papers and simulations and experiments, some industries are trying to incorporate and use such complicated systems. The best example may be the projected joint lunar base by China and Russia. In one of its pitch videos they showed a robot swarm on the moon together with a set of other different equipments. How or even whether it is really going to happen is still to be verified but at least the power of swarm robots is growing to a point where rocket scientists are considering to use this method. For the framework, there is really not much to be done, ROS is general purpose enough that it is hard to imagine a huge change would be needed on it so it can be used in the swarm robotics setting. One possible enhancement can be made is to further simplify its communication cost to support a wider range of hardware or support lower end hardware better. This is rather important as edge requires as little overhead as possible and cloud would favor those have a higher efficiency. There is no better way to reflect the fact where efficiency/time is money. After all, swarm robotics is a new and fast-growing category. And it is more or less apparent which cloud and edge would play an extremely significant role in this field.
Conclusion
This blog shares some stories on how cloud and edge can play a role in robotics field. Also, a more technical oriented blog featuring some interesting findings related to this will be shared at a later date.
“Hail the Omnissiah! He is the God in the Machine, the Source of All Knowledge.”
This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.