This year, the wave of Artificial Intelligence has swept the world again. At the beginning of October, the Nobel Prize in Physics was announced, and "machine learning expert" John J. Hopfield and "godfather of AI" Geoffrey E. Hinton won the prize together for their "fundamental discoveries and inventions in machine learning based on artificial neural networks". This event has sparked widespread discussion, not only because the Nobel Prize in Physics was first associated with AI technology, but also because neural networks and their associated deep learning theories have brought disruptive innovations to a wide range of fields.What is an artificial neural network?
An artificial neural network is amathematical model that mimics the way neurons work in the human brain, enabling it to recognise patterns and make predictions by learning from large amounts of data.
Image courtesy of the Internet Public DomainBy building multi-layer structures (i.e., deep learning), such networks can extract complex features from high-dimensional data, greatlyenhancing the capabilities of machine learning, especially in applications that deal with large-scale, complex data such as image recognition and natural language processing. End-to-end learning (End-to-End Learning) is an approach built on top of this deep learning framework, which centres on the use ofa single neural network model to generate the final output directly from the input data.End-to-End 'On-Road' Exploration
In autonomous driving, this architecture breaks the "bottleneck" of traditional modular systems.
Conventional autonomous driving systems rely on multiple independent modules, such as perception, localisation, decision-making and control, each of which is responsible for processing different types of information. However, in complex and changing road environments, such a modular architecture can lead to delays in information interaction, which in turn affects the speed of decision-making. 
Image derived from web-based public sources
In contrast, end-to-end learningtakes inputs and makes decisions directly from sensors (e.g., cameras, LIDAR, etc.)through a unifiedmodel. This architecture optimises the entire operation process as a whole, avoiding the bottleneck of information transfer in modular systems, and improving reaction speed and decision-making accuracy - in dynamic scenarios such as the sudden appearance of pedestrians or vehicle lane changes, the end-to-end model is able toadapt to the changesmore quicklyand respond accurately.
Image from the Internet public data
End-to-end "street" practice
Although the current end-to-end learning process is still difficult to explain and strongly relies on diverse data and other problems, but the significant advantages in dealing with complex tasks make it become the preferred method in many industries, such as body intelligence, driverless and so on. In the sanitation domain, the interaction mechanisms of obstacles are very complex, so end-to-end learning shows its unique potential.
In traditional sanitation work, whether it is manual or early automated equipment, it usuallyneeds to go through a complex process:firstly, the sensors collect environmental information, then the pre-processing module classifies and arranges the information;then the decision-making module makes a judgement on the basis of the collated information;finally, the actuator completes the corresponding action.
Although this step-by-step approach realises automation to a certain extent, the response is often slow in the face of emergencies, and there may be a delay in the transmission of information between the modules.
In contrast, end-to-end learning adopts a novel integration approach.
We apply this technology tounmanned cleaning robots, integrating the sensing, decision-making and execution processes through a unified neural network model to form an integrated and efficient working model that canbetter cope with complex and changing operating environments.Taking the unmanned cleaning robot Star Kyun® as an example, its sensors continuously collect data from the surrounding environment, such as road conditions, weather changes, traffic flow, and the dirtiness of the ground, and based on these real-time data, the end-to-end model can directly generate cleaning decisions, including travelling paths, obstacle avoidance strategies, as well as adjustments to the intensity and frequency of cleaning. This approach not only simplifies the decision chain, but also greatly reduces the delay phenomenon in information transmission, enabling the robot to respond to environmental changes quickly and almost senselessly. In addition, because all steps are completed within the same algorithmic framework, the robot can achieve more coordinated operation, thus improving overall work efficiency.Continuous"Evolution"of Robots
Robot scale is an important advantage of data pool accumulation.
KUSA Technology has accumulateddata from more than 60,000 sanitation vehicles. Relying on the powerful data processing capability of the cloud-based management system, KUSA Technologyprovides a rich data pool for end-to-end learning algorithm optimisation. In commercial operations in Shanghai, Jiangsu, Zhejiang and Sichuan, the daily sweeping data generated by StarHyun® is uploaded to the cloud in real time for centralised processing and analysis.
The system will extract key features such as different road conditions, weather, time periods, etc. from these data and continuously optimise the robot's algorithms so that it can respond to diverse operating scenarios more accurately and efficiently. Each new batch of data provides more comprehensive environmental feedback for the end-to-end model, thus enabling the robot to adapt more precisely to complex environments in long-term operation and truly evolve itself. Cloud-based management system (partial functionality)
Through this continuous learning mechanism, the unmanned cleaning robot Star Kyun® not only improves the efficiency and flexibility of cleaning operations, but also offers more possibilities for urban services.As technology continues toadvance and data continues to accumulate, end-to-endlearning will continue to advance thedevelopment ofunmanned sweeping technology, enabling smarter and more efficient cleaningoperations. In the future,with the expansion of more application scenarios, end-to-end learning is expected to play its unique advantages in more fields, further promoting urban construction and sustainabledevelopment.
