Autonomous AI at scale
Craig Davidenko 克雷格
President of Business Development at VTOL AEROSPACE? - UAV SYSTEMS Principal Urban Air Mobilities?
Building autonomous AI can be a daunting task that requires enormous amounts of technical ability sprinkled with lots and lots of datasets and the painstaking time needed to test the machine learning (ML) models, first in simulators and then out in the field.
Even when you have mastered all of that you still have to contend with updating the ML models on edge devices, for low latency, and processing and storing the roughly 12GB of data per second obtained about its environment from sensors such as LiDAR, cameras, radar, etc.
The problems are compounded further when you move from a single prototype to mass production as you now have to deal with managing these devices and ML models at scale.
To solve some of these problems, I have recently turned to two Internet of Things (IoT) technologies. The first is IOTA, which allows secure data transmissions to and from our sensors. IOTA is not based on a blockchain with its associated scalability challenges, energy consumption, and transaction fees. The IOTA tangle is a Directed Acyclic Graph (DAG) to create a distributed ledger that scales and has no transaction fees. Lightweight clients and the IOTA tangle’s off-line capabilities make it a perfect fit for our Internet of Things (IoT) sensor needs.
By enabling trusted connectivity and micro-transactions, IOTA uniquely enables the Internet of Things (IoT) for autonomous systems. Not only does this mean that we can now securely store and process all of the sensors data off the edge device but we can even buy and sell data directly on the device using feeless Machine-2-Machine (M2M) real-time micro-transactions.
Imagine a scenario where an Unmanned aerial vehicle (UAV) observes a new 20 story structure under construction in its flight path and automatically warns other autonomous UAVs or vertical take-off and landing (VTOL) aircraft in that area of the new structure for a small nominal fee that is negotiated directly between the machines themselves. These transactions are not just limited to UAVs or VTOL but can also come from smart cameras, self-driving cars or any IoT enabled device.
Currently, autonomous UAVs are very task or mission specific due to the way ML models and other algorithms are trained and built. An autonomous search & rescue UAV cannot be used to inspect bridges and power lines or vice verses This limitation is often the reason why completely autonomous UAVs has been slow to be adopted by the UAV industry. Who wants to own a UAV that can only do one thing when a piloted UAV can do multiple tasks/missions.
To address this issue and create an "AI on-demand" system I've turned to a blockchain based technology called SingularityNET. You can think of SingularityNET as being a sort of an “AI app store” where apps can not only serve users’ needs but invoke each other’s help to complete their tasks or projects. Any creator of AI can plug their algorithm into the Net, find customers and reap benefits from the network of other AI algorithms and is set to fundamentally change how autonomous AI is developed and deployed in the future. The AI on each UAV can now be switched in real-time to accommodate different task/missions thus wholly removing the single-use barrier to the adoption of autonomous UAVs.