The student team from Tsinghua University conducts testing for the Hitch Open World AI Championship on the Tianmen Mountain circuit in Zhangjiajie, Central China's Hunan Province on October 13, 2025. Photo: Courtesy of Tsinghua University

Intermittent signals, lack of visibility, hairpin turns, and plunging gradients - on the serpentine mountain road of Tianmen Mountain in Zhangjiajie, Central China's Hunan Province, known locally as the "Heaven-Linking Avenue," autonomous driving technology is being subjected to one of its most unforgiving real-world tests.

Far from the mist-shrouded cliffs, another milestone is unfolding in cities now that China has officially announced the country's first batch of market-entry approvals for Level 3 (L3) conditional autonomous driving vehicles. On Saturday, China's first official license plate dedicated to L3 autonomous driving vehicles was issued in Chongqing, marking an important shift that autonomous driving is moving beyond closed testing grounds toward real world usage.

Yet as industrial deployment accelerates, the question of safety, especially for edge cases, has only grown more pressing. 

Against this backdrop, a parallel experiment unfolding on Tianmen Mountain offers a revealing counterpoint. On the same stretch of road, widely regarded by industry insiders as "one of the world's most challenging AI testing grounds," a student team from Tsinghua University watched as their fully AI-controlled unmanned race car sped to victory at the Hitch Open World AI Racing Championship, setting a world record for autonomous vehicles tackling extreme mountain terrain. 

"High-difficulty 'right-of-way games' force algorithms to strike a balance between safety and efficiency," Sun Hang, chief engineer at the China Automotive Technology and Research Center, told the Global Times. "In a city of tunnels like Chongqing, frequent entry and exit put inertial navigation under constant scrutiny. These are exactly the metrics that L3 autonomous driving must be validated against under extreme conditions. In that sense, the Hitch Open competition aligns perfectly, and generates valuable experimental data."

The Tianmen Mountain road with 99 hairpin bends Photo: Courtesy of local tourism officials

An extreme test

On October 15, a peculiar race car pierced through thick fog on Tianmen Mountain. With no driver behind the wheel, it descended a vertical drop of 1,100 meters, threading its way through 99 perilous hairpin bends. The car belonged to a Tsinghua University student team and was controlled entirely by AI.

When it entered tunnel blind zones, multi-sensor fusion allowed it to "reorient itself" with precision. On slick downhill stretches, reinforcement learning algorithms kept the tires firmly planted. 

After 16 minutes and 10 seconds, it crossed the finish line, an unprecedented feat.

The team itself had been assembled hastily in late March. In just six months, they completed what amounted to a sprint from the laboratory to one of China's most treacherous mountain roads.

The Tianmen Mountain road spans 10.77 kilometers, coiling upward like a layered cake, hence its local nickname, the "Cake Road." With a maximum gradient of 14 degrees and a staggering vertical drop, it has earned the moniker "China's King of Extreme Roads." To the Hitch Open organizers and competing teams, it became known simply as the "AI Sky Road."

At the heart of the challenge is positioning, the AI car's sense of where it is. Without human judgment, the vehicle must continuously determine its precise position to match route planning, avoid hazards, and make accurate decisions.

During an initial test the team received a notification that their car stalled after just three kilometers due to map-loading lag, a fatal flaw with the 10-kilometer mountain race looming.

They used a "dynamic local map loading algorithm," inspired by the concept of assembling a puzzle with only those pieces needed. 

Team captain Qi Xiaojing told the Global Times that their goal was to design the fastest and safest path from mountaintop to base. A master's student, Qi had long hoped to bring laboratory research into the real world.

Ideals and reality

How close, then, are these extreme races to solving the real-world demands of L3 autonomous driving?

Sun emphasized that the Hitch Open Championship represents L4 to L5 autonomy, while most automakers remain focused on L2 to L3. "Because there are no passengers, extreme conditions can be tested without human safety risks," he said. "That's how you probe algorithmic limits."

Tang Minqin, global executive chairman of the championship, often invokes a famous engineering principle, saying "We chose Tianmen Mountain not because it is easy, but because it is hard."

All data generated during the competition, Tang told the Global Times, will be fully open-sourced. These rare extreme-scenario datasets, nearly impossible to collect on public roads, are invaluable for training physical intelligence models and advancing generative AI algorithms.

Qi revealed that following their debut race, the team has begun exploring collaborations with industry partners, aiming to translate competition-driven innovation into real-world applications.

Wang Junhao, president of the Shanghai Juneyao Group, the parent company of JuneYao-Auto, sees similar value. "This isn't just a contest between universities," he said. "It will generate new technologies and opportunities. The data alone is unquestionably usable."

Ultimately, experts agree that real, complex, and failure-prone extreme testing is an indispensable safety proving ground before autonomous driving can scale up. Its value lies not in spectacle or speed, but in rigorously defining system stability and risk boundaries.

According to Tang, future Hitch Open seasons will push further, introducing multi-vehicle overtaking races under unified obstacle-avoidance and vehicle-road coordination standards. And as Sun cautioned, L3 autonomy still demands careful evaluation of performance, liability, and risk governance. Approval is not permission for instant expansion, but a cautious beginning, under certain conditions.