What would you think if there was a puddle of water in front of you?

Maybe you'd instinctively walk around it, maybe you'd judge it to be a sign of last night's rain, or maybe, you'd ponder whether there was poor drainage nearby.

It's our instinct as human beings that our brains don't just see, but understand - building an internal model of the world from one experience to the next, a model that tells us: which things change, which phenomena have cause and effect, which decisions are worth caring about.

And what about robots?

In the past, they only recognised anomalies in images of water or radar reflections and avoided them according to the rules. They didn't "think", let alone understand. But now it is beginning to have "its own judgement".

Behind this is the "world model" at work. It gives the robot an imaginative view of the logic of its environment: not just seeing the current state, but predicting future changes, reasoning about causal paths, and anticipating potential risks. It knows that a leaf may be swept up by the wind, that a pedestrian may suddenly cross, that a stationary vehicle may be about to start. ......

But for such capabilities to be truly embedded in a robot's day-to-day decision-making, the key is not just to have a model, but also to have an architecture - the combination of a world model and an end-to-end system is becoming the next stage in the development of intelligent systems.

Most traditional robotic intelligent systems are modular in design: subsystems such as perception, localisation, planning, and control are independent of each other, requiring engineers to set hundreds or thousands of rules for each scenario. This type of system is prone to slow response and fragmented behaviour in the face of complex and unexpected cleaning environments in the city.According to industry experience,amodulardesignbasedself-drivingcleaning systemmayneed towritemore than 2,000 lines ofcode to cope witha singlecorner case,whereasa world model-driven end-to-end architecture can generate a large number of virtual scenarios in a simulator for training based on short-term real data, which makes the system more generalised and forward-looking.

End-to-end architecture for generalised training

Under such an architecture, perception, understanding and decision-making are no longer disconnected from each other, but as a unified neural network for overall learning and collaborative optimisation.The world model becomes the "brain" of the system: it receives continuous observations from sensors, predicts potential trajectories of the environment, and deduces optimal actions based on the target task.The whole process does not require an explicit definition of intermediate states, nor does it rely on static maps or hard-coded rules.

For example, in an urban open road sweeping scenario,based on a pre-trained model, a robot only needs to add tens of hours of real operational data to fine-tune a model that better meets the local user's sweeping habits, and reacts in advance to hundreds of thousands of changing conditions: it can judge that a certain vehicle may start temporarily and slow down its speed; it can understand the temporary parking situation on the side of the road, and flexibly adopt different conditions, such as not sweeping quickly or maximising the sweeping speed. It can understand the situation of temporary parking on the roadside, and flexibly adopt different strategies such as rapid passing without sweeping or maximising sweeping.

This kind of " understanding + evolution" ability is difficult for traditional systems to have.

Take the unmanned cleaning robot Star Kyun® as an example, each robot is equipped with a multimodal perception system, led by visual AI, combined with a variety of sensors working in concert,the massive amount of data captured by these sensors, after the edge of the edge of thedesensitisation, compression and otherpre-processing, uploaded to the cloud forself-labellingand training samples generated, and ultimately constructed for the learning of world models. "Spatio-Temporal Fragmentation.

Robots operating in complex urban scenes

In order to achieve efficient deployment, the system will use model distillation, pruning and quantisation and other technical means to compress large-scale world models into lightweight models that can be efficiently run at the terminal.These small models have faster inference speed and lower power consumption, and can independently complete real-time prediction and decision-making locally, thus realising the intelligent closed loop of "cloud-end collaboration".

Based on this, the robotbuilds a world model with generative prediction capability by combiningself-supervised learningandreinforcement learningmechanisms.Even in new environments, robots can simulate and generate future frames with the help of existing models, so as to evaluate the results of different decision-making paths in advance. This mechanism essentially previews current decisions in the "virtual future".

World model training process (based on temporal frame prediction)

For example, when there is an obstacle in front of the robot, instead of relying on predefined rules to make a judgement, the robot will use the world model to deduce: is the area passable? Is the obstacle stopped or in motion? Can an attempt be made to pass? In just a few hundred milliseconds, the system will complete the simulation internally and output the current optimal strategy.

Thanks to the intelligent cloud control platform, the world model also has a cross-device sharing mechanism.

For example, an unmanned cleaning robot learnt the pattern of "coping with the re-planning of paths due to the indiscriminate parking of shared bicycles" on the streets of Suzhou, and can instantly reproduce the same kind of early warning and adjustment capability in other regional devices by updating the parameters of the model.

As the complexity of sweeping scenarios continues to increase - including low light at night, seasonal leaf drop peaks, and areas of mixed pedestrian and vehicular traffic - the world model's learning capabilities continue to evolve.With the end-to-end architecture, each run of data is not just a trace of task execution, but also a collection of new knowledge. Through the closed loop of data, the system can automatically reflect and optimise the strategy after each operation, and eventually move towards a physical intelligence with high adaptability and long-lasting learning ability.

Currently, the world model has already played a role in a number of urban cleaning tasks.

When scenes such as waterlogging after rain, sudden pedestrian flow, and nighttime low light, which are difficult to be handled in advance by traditional systems, appear, the robot is able to complete the prediction and strategy switching within milliseconds, which significantly improves the traffic safety and sweeping efficiency. Especially in highly uncertain areas such as night-time intersections or around schools, the world model helps the equipment make more prudent choices by "previewing the future", avoiding many potential risks.

Robot operating in low-light environment

Meanwhile, the learning ability of the system is also "evolving" with the expansion of the operation area - the data from Chengdu's slippery roads in winter will provide stronger robustness for Shanghai's rainy season; the prediction experience of Suzhou's morning rush hour pedestrian-vehicle mixing has been synchronised and used in Hangzhou's core business district. The data from Chengdu's slippery roads in winter will provide more robustness for Shanghai's rainy season.Every data upload is a patch for "understanding the world"; every model update is an expansion of the robot's "behavioural boundaries".

In the future, with the continuous development of computing resources, algorithmic structure and sensing fusion technology, the end-to-end world model will no longer be limited to the recognition and decision-making stage optimisation, but will become the core architectural paradigm of the robotic system: it will no longer be a module, but the robot's consciousness hub.

Like human beings, robots can not only see what is in front of them, but also understand the world in which they live, and make reasonable, explainable judgements with cause-and-effect logic - this is exactly the significance of the existence of the end-to-end world model.