Author | Benyi Editor | De Xin
One day in 2023, Tao Ji walked into the W1 building of Changan Automobile. After 12 years at Baidu and a previous entrepreneurial venture, this independent automaker headquartered in Chongqing became his next destination.
In the autonomous driving circle, Tao Ji is a name that needs little introduction. He led the front-line team for the highly publicized Wuzhen autonomous driving fleet in 2016. For a period, there was a saying in the industry: "The outside world chases after autonomous driving experts, and the experts are all talking about where Tao Ji is going."
After his previous entrepreneurial venture ended, Tao Ji originally planned to dive into embodied AI startups and had even received investment intentions. During the same period, several automakers extended olive branches, but Changan ultimately won him over. Changan has a very pragmatic ethos. During discussions with Changan's senior management, a few words left a deep impression on Tao Ji: "Focus on autonomous driving first, and robots later; I will definitely support you."
Getting hands-on and prioritizing mass production, this attitude later became the core consideration for Tao Ji's R&D decision-making at Changan.
Three years later, the end-to-end advanced driving assistance system independently developed by the team led by Tao Ji entered mass production. Two hours after the pre-sale of the Qiyuan Q06 opened, orders exceeded 11,000 units. This model is equipped with Tianshu Pilot Ultra and Tianshu Pilot Pro, adopting Horizon Robotics' J6P and J6M computing platforms respectively. The software algorithms both feature Changan's in-house developed single-stage end-to-end system.
About a year and a half ago, HiEV experienced Changan's in-house highway NOA (Navigation on Autopilot) solution at the time. This time, following the Q06 pre-sale launch, we also experienced the urban NOA of the Tianshu Pilot Ultra version in Chongqing, and our subjective experience refreshed our understanding. Although the model version tested that day was relatively conservative, the strong safety boundaries and high system ceiling were clearly felt.
If we exclude the three new EV forces, it can even be said that this is currently the best in-house developed solution among independent automakers.
Over the past few years, many domestic automakers have shouted the slogan of "in-house development," but few have actually delivered. "Today, the number of automakers and suppliers that have truly mass-produced single-stage end-to-end systems can be counted on one hand," said Tao Ji.
After three years of focused R&D in Chongqing, Changan has taken a crucial step in mastering core intelligent technologies.
Others Withdraw, Changan Quietly Catches Up
Actually, Changan has never lacked opportunities to access mature autonomous driving solutions.
Looking back, Changan's exploration of intelligence can be divided into four stages. From 2016 to 2020 was the germination period, starting with internal in-house exploration and external collaboration; the autonomous driving module for Deepal was provided by a joint venture between Horizon Robotics and Continental. From 2021 to the first half of 2023, the company organized multiple parallel lines internally to explore cutting-edge R&D in autonomous driving. For a period, what attracted the most external attention was that with the founding of Avatr, Huawei brought a full-stack intelligent cockpit and autonomous driving solution, including urban NOA with LiDAR.
The cooperation between Avatr and Huawei was of extraordinary significance to Changan. Through this partnership, Changan also saw the entire chain for the implementation of high-level autonomous driving, such as the underlying E/E (Electrical/Electronic) architecture, systematic AI capabilities, and the scale of talent and strategic resources required.
Changan Automobile has a history of over 100 years, starting from the Shanghai Yangpao Bureau. It relocated from Shanghai to Nanjing and then to Chongqing, transitioning from military to civilian use, and from micro-vehicles to passenger cars. Historical imprints have shaped Changan into an automaker with a strong sense of crisis; mastering its own destiny and core technologies is almost ingrained in its bones. Therefore, after 2023, as industrial technology and timing matured, Changan's in-house autonomous driving R&D accelerated once again.
In a media interview, Tao Ji used OTA (Over-The-Air) as an example to discuss the benefits of in-house development. Suppliers need to serve multiple automakers, and OTA must align with major version release schedules. If Changan chooses in-house development, it can improve faster based on user feedback. His thinking also extends beyond autonomous driving; in the future, the cockpit, chassis, and driving assistance need closer coordination, and multiple types of data may need to be integrated during the training phase.
For OEMs to participate in product definition, they must first master the capabilities of processing data, training models, and validating results. The attitude of Changan's management is also very clear. Chairman Zhu Huarong characterized the CNY 10 billion-level investment as "one of the largest strategic investment projects in Changan's history." Executive Vice President Yang Dayong was even more direct: "Changan brand models will be equipped with, and will only be equipped with, Changan's in-house developed Tianshu Pilot."
In-house development is not an alternative; it is the only option.
In the second half of 2023, Changan officially switched to building its in-house R&D system. Over the past decade, Changan's cumulative investment in intelligence (including autonomous driving and intelligent cockpits) has exceeded CNY 60 billion, with relevant R&D teams comprising over 7,500 people. The end-to-end project later received a special budget of CNY 10 billion approved by senior management.
In many exchanges, reflecting on his journey from Baidu's autonomous vehicles to entrepreneurship, he could not hide his regret that commercialization was not fully realized. At Changan, he had the opportunity to turn technology into products that people actually buy and use.
In-House End-to-End: "Build the Kitchen Before Cooking"
When Tao Ji first joined Changan, the autonomous driving team had several hundred people, but only about thirty were working on data infrastructure and development tools; they didn't even have an in-house BEV (Bird's Eye View) perception solution yet.
Changan's past accumulation in autonomous driving R&D was more focused on planning, control, and engineering delivery. Engineers had experience in how vehicles execute commands; however, continuously producing training data to teach models how to drive required a different organizational logic.
Tao Ji did not rush to advance algorithm R&D. Instead, he launched a project called the "AI Cornerstone Project," even giving it an English name to make it more appealing. He then prioritized assigning campus recruits to infrastructure work, building the data closed-loop, training platforms, and simulation systems one by one.
In an industry atmosphere where everyone is eager to roll out demos, spending a whole year on tasks with no visible direct output requires resolve. To motivate the internal team, he used the analogy: "Infrastructure is the kitchen, and algorithms are the dishes. Without a good kitchen, even the best chef cannot cook a meal."
After joining, Tao Ji also pushed for another initiative: establishing a horizontal evaluation system. Scenarios were divided into four categories, using the standards of experienced drivers as baseline references, and all model versions had to pass this test. We can get a glimpse of the scale of this "kitchen" from a few sets of numbers.
500 PB of storage capacity, a total data pool of 1 million hours, over 300 sets of models trained in parallel, and over 100 sets of new models iterating simultaneously. 4DGS scene reconstruction is used for simulation, with more than 15,000 core scenarios forming the test library, and models are ranked daily. In Dianjiang County, Chongqing, Changan also built the world's first digital-physical integrated unmanned testing ground.
Tao Ji considers himself satisfied with this infrastructure "kitchen," stating, "It is already in the first tier of the industry."
Currently, autonomous driving systems in the industry based on single-stage end-to-end architectures are gradually entering a plateau in performance growth, which is the root cause of the current numerous debates over new routes such as VLA (Vision-Language-Action) and WAM. However, Changan's R&D team told us that internal evaluations based on their designed architecture and data closed-loop still show a very steep growth curve.
Besides the data system, talent organization is the core of in-house autonomous driving.
Tao Ji summarizes the talent organization as a 1:4:N ratio: one leader capable of judging the direction, paired with several core members with practical experience, who then drive the execution team. People from Baidu, Huawei, internet backgrounds, and traditional automakers are brought together; there were cultural collisions in the early stages, which turned into complementary strengths after integration.
During a small-scale media communication after the Qiyuan Q06 pre-sale launch event, several autonomous driving R&D leaders brought in by Tao Ji came from diverse backgrounds.
The head of the data platform and toolchain is a veteran Changan employee who has worked at Changan for 16 years; the head of end-to-end algorithms comes from Baidu's visual large model team; and the head of autonomous driving products has experience at new EV forces like NIO and Li Auto. Taking the data platform team as an example, although the head is a veteran Changan employee, the core leaders reporting directly to her are all from external backgrounds to supplement capabilities in data platforms, simulation, and cloud models.
After the organization took its initial shape, the team first used the method of perception large models combined with hand-written planning and control to develop highway NOA, while simultaneously supplementing the toolchain. By the second half of 2024, when the infrastructure had been built for about a year and the "kitchen" was basically in place, Tao Ji believed that continuing to use the old methods to develop urban NOA was no longer appropriate and began pushing the end-to-end project.
Choosing Single-Stage, Taking No Shortcuts
There are two approaches to end-to-end. The two-stage approach trains perception and planning/control separately and then splices them together, which is easier to debug in engineering. The single-stage approach connects everything from sensor input to driving behavior output in one model; it is harder to train but has a higher ceiling.
Changan has run both routes, conducting parallel experiments for three to four months. The two-stage approach required supplementing urban perception data and high-precision annotations, as Changan's previous reserves for highways were insufficient. The single-stage approach reduces intermediate processing steps but has higher requirements for simulation and validation. Both sides had a massive amount of work to catch up on.
By the end of the first quarter of 2025, the team decided to go all-in on the single-stage approach. Tao Ji's exact words were, "Reaching 70 points has theories to refer to, but achieving truly good performance really relies on experimental science." There are no shortcuts, only trial and error version by version.
The training pipeline is a funnel. Pre-training, supervised fine-tuning, reinforcement learning, and fallback intervention allow problems to converge layer by layer. The team internally summarized this methodology into two words: "Probabilistic Intelligence + Deterministic Delivery." The performance of AI models is inherently probabilistic, yet the automotive industry requires deterministic safety standards. The tension between these two runs through the entire process of mass-produced autonomous driving.
This methodology serves a specific task: when a problem occurs on the road, how to fix it in the next version. Testers send back problem clips, the cloud analyzes the scenarios and tags them, and then finds high-quality driving samples in similar situations from existing data. We first filter and proportion the data, then use different data combinations to train different models; subsequently, we reconstruct scenarios using technologies like 4DGS and compare them through closed-loop simulation to select models. The biggest challenge in the early stage was lateral displacement: at that time, it was basically only possible to handle scenarios with longitudinal movement and no lateral movement. Now this problem has been gradually overcome, and scenarios with significant lateral changes can also be used for closed-loop simulation, so we have achieved scaled application in model selection.
However, the data itself can also bring problems. Tao Ji mentioned in a previous interview that requiring data collection drivers to strictly follow rules might result in data that is too slow; switching to ride-hailing data improves efficiency, but then requires clearing clips of crossing lines, erratic lane changes, or even unprompted stops. Which behaviors the model learns depends on what data the team feeds it. The head of Changan's autonomous driving algorithms believes the core is to have a robust model evaluation mechanism—without comprehensive model selection, a version the team thinks is good might not be recognized when it reaches users.
During the large-scale road testing phase, Tao Ji introduced that 590 volunteers were distributed across 45 cities, accumulating over 1 million kilometers of real-road testing. Daily war-zone meetings were held to align on issues and iteration pace. Models are ranked daily, and every version must pass the horizontal evaluation system.
Tao Ji described this period as "changing the wheels while driving." The team maintained the mass-production system already running while simultaneously stacking the end-to-end architecture on top.
In October 2025, the team saw key results in single-stage end-to-end R&D and judged there was an opportunity to catch up with the mass production schedule of the Qiyuan Q06 the following year. It was not until March 2026 that the version was basically confirmed to match the vehicle's launch timeline.
However, reaching the deadline does not mean the version is automatically released. Tao Ji listed several requirements, including horizontal evaluation scores, takeover-related statistical indicators, and quality defects. One of them is a requirement familiar to automotive engineers: zeroing out S-class issues.
This requirement caused a dispute between the two teams. The algorithm team believed that statistical indicators had improved and the proportion of issues had decreased, so the version could be released. The quality team disagreed, stating that severe and unacceptable issues remained unresolved. The algorithm team's reasoning was that AI systems have inherent uncertainty, and it is impossible to demand all issues be zeroed out. The quality team's concern was that a single issue falling on a specific user constitutes a complete failure experience.
Later, both sides calmed down and re-examined the specific issues pointed out by the quality team, finding that some were not extreme sudden situations and would not be considered difficult by human drivers. Given this, it was hard to leave these issues to users solely on the grounds of improved average indicators. The quality team also adjusted the counting method, introducing statistical metrics such as per-unit mileage, because continuous testing will continuously discover new issues, and the total number of defects might accumulate higher and higher; they cannot judge whether the software has improved solely by staring at the cumulative quantity.
Tao Ji's team mentioned that they will definitely rush into the first tier this year. Everything about the product prioritizes safety first. Media test drive experiences reflect this; for example, when encountering a bicycle ahead on a mountain road, the test vehicle slowed down and followed, without crossing the solid line to bypass it, trading off for a safety margin.
"I Am Not Easily Tempted"
The two hottest concepts in the industry over the past two years, VLA and world models, are viewed very calmly by Tao Ji.
Will the solution in hand become obsolete just as it is developed? Tao Ji's criterion for judgment is to see "what problems the new technology solves."
Tao Ji stated that small internal teams at Changan are testing the VLA technical route, with the conclusion that it shows potential in obstacle avoidance timing, lane selection, and dynamic interaction, but there is a gap in the stability of longitudinal and lateral control and speed rationality compared to imitation learning solutions. Tao Ji's attitude is that before VLA is proven to comprehensively surpass end-to-end, they will not bet on switching.
"I am not a person who is easily tempted," Tao Ji said. He also emphasized that the voice-based lane changes and memory preferences already demonstrated by Tianshu Pilot are interactive capabilities achieved by integrating vision-language models with the autonomous driving system, which should be viewed separately from the technical route of the driving model itself.
Frankly speaking, today's algorithm testing methods are still relatively primitive, mainly relying on open-loop simulation supplemented by real-vehicle testing. However, at the Qiyuan Q06 launch event, Tao Ji already showcased new progress in scene reconstruction and generalization testing, with simulation helping to filter models.
His judgment differs from the mainstream discussion direction. In Tao Ji's view, what current autonomous driving AI lacks most is a "mental model."
Whether end-to-end or VLA, current AI lacks the ability to understand the intentions of other traffic participants. Human drivers can judge from the subtle movements of another vehicle whether it intends to change lanes or is hesitating, and whether the delivery vehicle ahead is pulling over or continuing. This mind-reading ability is the core of primate intelligence and a missing link in the current AI evolution chain. Tao Ji believes that solving this problem takes higher priority than chasing VLA or world models.
Tao Ji summarizes the future technological landscape as "Action AI," where world models, ontological models (the vehicle's own physical models), and mental models co-evolve. A further direction is controlled RSI (Recursive Self-Improvement), allowing models to iterate autonomously in a sandboxed cloud environment, with humans only setting boundary conditions.
Regarding the current single-stage model, Tao Ji believes "there is still room for improvement." The current training data consists of about 20 million driving clips, corresponding to 160,000 hours, and the total amount of cleaned and labeled data is about 1 million hours. Han Junyu also pointed out that some hesitation and experience issues discovered during nationwide generalization can still be improved through targeted data processing.
However, doing mass production at an OEM brings Tao Ji different troubles compared to his previous engineering implementation experience.
He mentioned in a previous interview that the team wanted to make high-level autonomous driving the best, but after finishing it, found the platform too expensive and the vehicle models unwilling to adopt it. Changan's products have a wide price range, and R&D investments ultimately need to be amortized through actual deployment. Therefore, at this stage, he values the ability to prune, quantize, and deploy models across different chips more.
Currently, Tianshu Pilot Ultra is equipped with Horizon Robotics' Journey 6P chip, with a single-chip computing power of 560 TOPS, and the algorithm and vehicle integration are completed by Changan itself. Through a unified operator library, the in-house WBE inference framework, and a model pruning and quantization toolchain, the same set of algorithms can be deployed on different platforms such as J6M, J6P, and Thor, without being locked into a single supplier.
"We have many tricks that haven't been deployed yet," Tao Ji said during media exchanges.
Currently, the data used for end-to-end training (20 million clips, about 160,000 hours) accounts for only one-sixth of the cleaned and available data.
In March this year, Changan obtained an L4 Robotaxi test license, and a mutual data feeding loop is forming between this line and mass-production autonomous driving.
Changan's current approach is to build a universal architecture at the group level, where computing power and sensors can be flexibly plugged and unplugged according to brand and model positioning. Qiyuan has achieved stable profitability in the market below CNY 150,000. After the launch of the Q06, Qiyuan's average monthly sales have the opportunity to reach about 60,000 units, with a total annual sales volume of around 400,000 units. This means the Q06 needs to contribute about 15,000 units per month and must penetrate the CNY 150,000 to CNY 200,000 price band, which Changan has never successfully tapped before.
He Gang, Vice President of Changan Automobile and President of Changan Technology, predicts the ultimate outcome of the industry: the autonomous driving industry may complete its consolidation around 2028, ultimately leaving only five or six solution providers. The error-tolerance window for Tianshu Pilot is not infinite.
Going forward, Changan needs to use the different configurations chosen by users and continuously iterated versions to prove that this in-house R&D system can operate in the long term. These unplayed cards will be the key variables moving forward.