A single self-driving car making a run is one thing. Running thousands of them, every day, is another. Hello Autonomous is applying its base of more than 340 million carpool users and years of mobility operations to Robotaxi, connecting dispatch, charging, maintenance, and monitoring so the service can run reliably every day.

By Sang Min Han _ han@autoelectronics.co.kr
한글로보기





How do you take validated autonomous-driving technology and turn it into a service that dispatches cars, charges them, maintains them, and carries passengers every single day? Liu Xiaoxi, vice president of Hello Autonomous, says the driving algorithm alone doesn't answer that question.
Take the driver out, and users still need to get where they're going, safely, on their own schedule. A human driver never just drives. They check on the car, notice small problems, talk to passengers, handle whatever comes up. In a driverless vehicle, all of that judgment and action falls to the car, the cloud, and the control center, and the more advanced the driving system gets, the more operational responsibility the operator has to carry.
Liu described Hello as a platform running bike-sharing, battery swapping, carpooling, and car rental together, covering trips ranging from a few kilometers to several hundred kilometers. It grew out of bike-sharing into a broad mobility platform spanning shared two-wheelers, battery swaps for riders, carpooling, and car rental — different kinds of trips, connected on one platform. It's not a carmaker, and it didn't come up through the traditional autonomous-driving startup path.
Hello says its carpool service has served more than 340 million users, and that user base and years of running the operation are what Hello points to as its case for entering Robotaxi — not a patent portfolio, not a fleet of test vehicles.
Removing the driver doesn't create a new mobility business on its own. Dispatch, charging, cleaning, maintenance, accident response, billing, and customer service all have to run every day without fail. Someone has to decide which neighborhoods need cars during the morning rush, when to pull a car in to charge, whether a vehicle showing early signs of trouble keeps running or goes to the shop. Operating a large fleet every day requires capabilities that go well beyond making an individual vehicle drive itself.
To extend that operational experience into autonomous driving, Hello signed a strategic partnership with CATL and Ant Group in April 2025. In June, the three companies' investment vehicle jointly founded Shanghai Zaofu Intelligent Technology. Company registration records list the firm's registered capital of RMB 1.288 billion; the companies have said combined first-phase capital and investment exceeds RMB 3 billion. Alibaba joined the partnership through cloud computing and its Qwen large language model.
Zaofu was a legendary charioteer from ancient China, known for his skill at handling horses — a name meant to signal mastery of the vehicle. Hello contributes its user base and operating experience; CATL supplies batteries and the energy ecosystem; Ant Group handles payments and digital services; Alibaba provides cloud and large-model support. The structure treats autonomous driving less as a single product than as an industry that links vehicles, energy, payments, AI, and operations.
Coming late to the field shaped Hello's technical choices as much as anything else. Instead of retracing the modular path earlier autonomous-driving companies built up over years, Hello started with an end-to-end architecture from day one. Traditional systems split sensing, perception, prediction, decision-making, planning, and control into separate modules — a well-proven approach, but one where added modules mean added complexity, and an error in one stage can carry through to the next. Hello instead feeds camera, lidar, millimeter-wave radar, and SD-map data into a single closed loop covering feature extraction, vectorization, a diffusion model, and decision and planning together, learning vehicle behavior directly from real driving data rather than combining fixed rules stage by stage.




CATL, Ant Group, Alibaba, and Hello formed a strategic partnership around Robotaxi in 2025. Zaofu Intelligent Technology, established through an investment vehicle backed by Hello, CATL, and Ant Group, plans to invest more than RMB 3 billion in its first phase in Level 4 autonomous-driving R&D and commercialization. Alibaba supports the partnership through cloud computing and its Qwen large language model.



Loading an end-to-end model onto cars running every day

Liu said end-to-end development requires large-scale data, large-scale compute, and large models. Hello is building a system in which vehicles and the cloud continually exchange data and model updates.
Hello's plan calls for more than 10 million usable data clips by 2026, ready for production development. The company pulls driving behavior from its mobility services, automatically identifies and refines the clips with the highest training value, then runs them through AI pre-screening and human review before they become training data. Hello focuses on quickly identifying the situations that provide the most useful training data: an intervention, an unpredictable move by another vehicle or pedestrian, a complicated intersection, or a construction zone.
To support that, Hello is building a large-scale GPU computing cluster on Alibaba Cloud. With the Qwen team, it's co-developing a large model for data mining, world-model-based simulation, and a closed-loop Vision-Language-Action (VLA) model. The car generates data; the cloud filters it and trains the model; the improved model ships back down to the car. When real-world driving turns up a new problem, the cycle runs again.
Where Leapmotor turns driving data into a validation asset reused across countries and vehicle programs, Hello uses a continuous cycle of data selection, training, and deployment to improve the model more quickly. Both companies emphasize the same thing: not the volume of raw data, but the ability to refine it into something reusable.
An AI model alone doesn't make a Robotaxi. The algorithm has to run on a vehicle built to survive long hours of high-frequency commercial service.



Hello Autonomous chose an integrated end-to-end architecture over the traditional Robotaxi approach of splitting sensing, prediction, decision-making, planning, and control into separate modules. It feeds camera, lidar, millimeter-wave radar, and SD/LD map data into a single pipeline covering feature extraction, vectorization, a diffusion model, and decision-making and planning.



Robotaxi HR1

Hello's first Robotaxi, HR1 (Hello Robot 1), was built on the Dongfeng Venucia platform, designed from the start around Robotaxi's vehicle specifications and production requirements rather than retrofitted with sensors and a computer after the fact. HR1 carries redundant design across six areas: chassis, communication, power, sensing, compute, and positioning. If the primary braking system fails, a secondary system brings the car to a stop. A dual-SIM, dual-link 5G TBOX keeps the car connected through two separate carrier networks and the cloud. The 12V main power and backup power run on separate circuits, and camera, lidar, and millimeter-wave radar sensing overlap to cover the surroundings redundantly.
The primary and backup compute units both run on automotive-grade domain controllers, with combined performance above 2,200 TOPS. If one compute unit or one network fails, the car can still drive safely or come to a stop.
Boarding and alighting are part of operational safety too. The moment a passenger opens the door, a bicycle, e-bike, motorcycle, or pedestrian could be approaching — a risk the operator has to manage even when the driving system itself never caused a collision. HR1 places separate sensors around the vehicle specifically to catch that kind of close-range approach.
Hello ran HR1 through six categories of production-vehicle validation: environment and climate, durability and reliability, E/E architecture, EMC, sealing and protection, and performance and safety. Testing covers conditions from -40°C to 85°C, high-frequency operation, and repeated door cycling. On top of redundant power and communication, millisecond-level fault diagnosis and isolation meet ASIL-D functional-safety requirements.
The purpose of this testing goes beyond completing a single successful run. It is meant to confirm that the vehicle can maintain the same level of performance and safety through long hours of repeated operation, carrying passenger after passenger.



HR1, Hello's first Robotaxi, was built on the Dongfeng Venucia platform. It carries redundant design across six areas — sensing, compute, positioning, communication, power, and chassis — with the primary and backup compute systems delivering combined performance above 2,200 TOPS.



An operating system for sending cars out every day

Liu said field operators and safety personnel need real-time visibility into every vehicle's status, that an accident has to trigger an immediate check of records and data, and that dispatch planning and operational efficiency all have to run through one system.
Hello puts more emphasis on this operating platform than on the driving algorithm itself. A Robotaxi operator has to know the status of every vehicle in real time: which car goes out when, when it charges, when it needs service. When something goes wrong or an accident happens mid-route, the operator has to pull vehicle status and driving records immediately and preserve the relevant data.
Remote safety personnel fall under the same system. How many vehicles can one operator monitor? Which alerts get handled first? After a remote intervention, does the car keep running or get pulled back in? Utilization, downtime, charging time, service frequency, complaints, profitability — all of it has to connect in one place before any of this becomes a business.
This work sits outside the driving algorithm, but it's central to commercialization. A car that can technically drive itself still makes no money if it sits idle because charging, maintenance, dispatch, and monitoring aren't connected. Building an autonomous vehicle and running a large fleet of them every day are different capabilities, and Hello frames Robotaxi commercialization as resting on both: the end-to-end model and the redundant, purpose-built vehicle provide the technical basis for safe commercial operation, while dispatch, monitoring, charging, maintenance, and safety management are what get that car out on the road every day.



The cockpit AI agent

A driverless car can become a new kind of service space. Passengers can watch video or listen to music en route, or order food or coffee timed to their arrival. The link is a cockpit AI agent that reads a passenger's situation and intent and carries out the service — say a passenger mentions wanting coffee, the agent checks the destination and arrival time, finds a suitable shop, places and pays for the order, and times it so the drink is ready when the passenger arrives. Connect fare-based transport to services beyond the ride itself, and a driverless car becomes a platform for services well beyond transportation. This remains a direction Hello has laid out rather than a finished business model.
HR1's actual operating scale, safety record, utilization rate, frequency of remote intervention, and per-vehicle cost and profitability are the metrics still to be confirmed. The technology and investment Hello has presented so far don't settle whether large-scale commercialization will succeed. Still, Liu's presentation pointed to a shift underway in Robotaxi competition, from whether an autonomous vehicle can drive to whether an operator can send thousands of vehicles out safely every day, match them to real demand, and pull a broken-down car in and get it back on the road.
The real test begins after the algorithm reaches the vehicle: whether that vehicle can finish today's service and return to the road tomorrow.

AEM(오토모티브일렉트로닉스매거진)