Find Sites

From high-precision maps to universal intuition: how Zelos Technology reshapes the underlying logic of autonomous logistics

For a long time in the past, the autonomous driving industry was firmly locked in by a seemingly natural premise - to make a car run safely in a city, it must first draw a high-precision map with centimeter level accuracy. This map is like a detailed road book, recording the curvature of every lane, every traffic sign, and every ramp. After vehicles hit the road, they can follow the road book. However, reality is cruel: urban roads change every day, construction fences are erected today and dismantled tomorrow, temporary traffic lights suddenly appear at a certain intersection, and newly built branches have not even been included in navigation software. The update speed of high-precision maps can never catch up with the pace of urban growth, and what's even more deadly is that once a car drives out of that carefully mapped "comfort zone", it immediately becomes a blind person who can't move a step. This is exactly the old industry order that Zelos Technology is attempting to fundamentally overthrow.
At the beginning of its establishment in 2021, Jiushi Intelligence chose a more difficult path: instead of constantly supplementing vehicles with increasingly thick maps, it is better to let it learn to use its eyes to see the road and its brain to make judgments like humans. The core breakthrough of Zelos Technology is first reflected in the so-called "graph free" path. The traditional driverless system follows the two-stage logic of "building maps first, then driving", while Zelos completely abandoned this mode and relied on real-time pure vision and multimodal perception to complete environmental understanding. This means that vehicles no longer need to know in advance where they are going. As long as there is a road ahead, they can use real-time visual information collected by sensors such as cameras and LiDAR to dynamically recognize lanes, pedestrians, vehicles, and obstacles, and make driving decisions based on this. The transition from "recognizing the road" to "seeing the road" is essentially the evolution of autonomous driving from a specialized skill defined by maps to a universal intuition with generalization ability. For the high-frequency scenario of urban distribution logistics, the value of this capability cannot be overemphasized - express delivery addresses change every day, internal roads in industrial parks are often not on any commercial map, and unnamed paths at the junction of urban and suburban areas are even more common. Zelos Technology allows vehicles to "drive wherever there is a road" and no longer be restricted by limited surveying ranges.
Having only perceptual ability is far from enough, as driving is never a purely visual recognition problem, but a complex decision-making process filled with common sense judgment and social interaction. An experienced driver will instinctively slow down when they see a child chasing a ball on the roadside; If you see a vehicle ahead turning on its turn signal and pulling over to the side, you may anticipate that it may require people to be driven; When I see an old man on a tricycle hesitating at the intersection, I will patiently wait instead of honking to urge him. These subtle judgments based on common sense of life are precisely the biggest shortcomings of autonomous driving in the past decade. To make up for this lesson, Zelos Technology introduced the VLM/VLA large model architecture, which directly embeds the visual language model and visual action model into the intelligent decision-making chain of the vehicle. The multimodal understanding ability of large models enables vehicles to recognize complex business scenarios where corresponding terms cannot be found in standard traffic rules, such as temporary stacking of goods at delivery points, security guards waving to give way in parks, and sudden crowds of people coming out of school gates. It is no longer just a code system that mechanically reacts according to preset rules, but has begun to possess some kind of "cognition and common sense" that is close to human drivers, capable of predicting potential traffic risks and outputting higher-order driving strategies.
Zelos Technology also builds an end-to-end intelligent control link between perception and decision-making. The traditional autonomous driving assembly line is highly modular - the perception module outputs a list of obstacles, the prediction module infers obstacle trajectories, the planning module generates paths, and the control module translates the paths into actions of the steering wheel and throttle. There is information loss and error accumulation between each link. Zelos' approach is to vectorize the perceived obstacles and map features into the BEV features of the bird's-eye view, allowing the environment, vehicles, pedestrians, and roads to be jointly modeled in the same high-dimensional feature space. This integrated architecture enables the system to synchronously predict the motion of surrounding dynamic participants and plan the output of the main vehicle's driving trajectory within a millisecond time window, rather than sequentially observing, thinking, and acting step by step. The result is that the vehicle's driving trajectory becomes more natural and smooth, with acceleration and deceleration no longer stiff and jerky like traditional unmanned vehicles. The actions of following, turning, and avoiding are closer to the driving style of human experienced drivers, which is crucial for improving the acceptance and sense of safety of other participants on the road.
What supports the operation of all these algorithms is a high-performance automotive grade computing unit that has also undergone full stack self-developed. In the autonomous driving industry, the contradiction between computing power and power consumption has long existed - to run complex perception and large model inference, powerful computing power is required, but the space and power in the car are extremely limited. Excessive power consumption not only squeezes the available power of the cargo box, but also brings engineering problems such as heat dissipation and reliability. The answer given by Zelos Technology is a highly integrated L4 autonomous driving computing architecture, based on ARM architecture to achieve ultra-low power consumption of less than 100 watts, while achieving global leading performance, capable of supporting collaborative inference of timing occupied networks and end-to-end models. This kind of engineering capability that maximizes computing power within a limited power budget is the implicit threshold for unmanned driving to move from laboratory demonstrations to large-scale production operations. It is also one of the underlying guarantees for Jiushi Intelligence to deploy unmanned delivery vehicles to over 100 cities worldwide and achieve a fleet size of tens of thousands of vehicles.
Looking back, Zelos Technology represents not just a breakthrough in a single point algorithm, but a complete technological paradigm shift from hardware computing, perception architecture, decision models to vehicle engineering. It marks the departure of the autonomous driving industry from the era of "precision farming" based on high-precision maps, and instead moves towards a stage of "general intelligence" driven by generalized perception and large model cognition. When an unmanned delivery vehicle can smoothly navigate unfamiliar roads without any pre surveying, understand the human driver's gestures and gaze that can only be understood, and complete every turn and avoidance with near human fluency, the day when unmanned driving truly enters the daily logistics scene on a large scale will no longer be far away. The significance of Zelos Technology lies in its practical engineering applications that are implementable, mass-produced, and scalable, proving that this more difficult but also correct path is feasible.