The steering wheel on an Inceptio Technology autonomous truck in Jinan, Shandong Province, China, on Thursday, April 18, 2024.
Bloomberg | Bloomberg | Getty Images
BEIJING — While AI updates make headlines every few weeks, those advances are not enough to get self-driving vehicles on the road more quickly.
That’s according to Chinese autonomous trucking companies, who say that improvements in large language model, from Anthropic’s Claude to China’s DeepSeek, have little impact on the timeline for vehicle deployment.
“The world’s best linguistics [expert] doesn’t mean he’s a good driver,” Pony.ai CEO James Peng told reporters last week. “AI is a very broad term. They’re completely different things. Absolutely … zero relevance.”
“When we process language, when we play sports, when we drive we all use different skills,” he said.
Autonomous driving uses artificial intelligence to imitate a human driver with a combination of sensors, chips and algorithms. But the real-world training data needed is very different from what powers large language models such as OpenAI’s ChatGPT, requiring what’s called world models.
Inceptio, a self-driving truck startup, is still sticking to its timeline for a mid-2028 commercialization milestone, unaffected by the broad advances in AI, CEO Julian Ma told CNBC.
By the third or fourth quarter of 2028, he expects Inceptio will have racked up 5 billion kilometers (3.1 billion miles) of truck driving data in China — enough to allow fully autonomous heavy-duty trucks to ply public roads.
With 5 billion kilometers in collected driving data, AI can extrapolate that into 50 billion km of experience in a world model — which is then sufficient to allow a heavy-duty truck to drive completely on its own, Ma said. He expects the trucks can then start operating without any humans inside in certain parts of the country.
Achieving that goal in about two years is already quite fast, he said, noting that in order for driverless trucks to become a widespread reality, they will need partnerships with manufacturers and regulatory approval — in addition to the tech.
Autonomous vehicles rely heavily on data about driving on roads. Just like robotaxi companies, self-driving truck operators run manned tests in order to gather training data safely.
Inceptio has by far recorded the most commercial autonomous truck miles in the industry, exceeding U.S. rivals, according to ARK Invest’s Big Ideas 2026 report in January. At the time, the company had driven 250 million miles — exponentially more than fellow Chinese autonomous driving company Pony.ai, which held second place at 4.2 million miles.
U.S.-based rivals Aurora, Kodiak and Gatik rounded out the top five, with a combined 8.9 million miles, according to the report.
Inceptio’s Ma said in late April the company’s trucks had driven 700 million kilometers (434.96 million miles), and aimed for 1 billion kilometers (621.4 million miles) by the end of the year. He said the company can use AI to identify which specific scenarios to focus on for gathering test data.
At the Beijing auto show, Pony.ai also announced an upgrade to its PonyWorld 2.0 AI model to improve its ability to collect specific data and train the model more efficiently. The company, which already operates robotaxis in China and other countries, unveiled a fully driverless light-duty truck that it developed with battery giant CATL.
Regulatory challenges
While China has 5-year development plans that increasingly emphasize tech goals, Ma said it’s often companies that take the lead in driving innovation.
“We make it happen,” he said, before regulators see the technology in action and are convinced enough to provide policy support.
But it’s clear that there’s a long way to go before you see trucks and cars running around the country without drivers.
“Automobiles are actually the most challenging area for AI, and exceeds the difficulty of embodied AI to some extent, because it involves safety,” Ma said. Embodied AI includes humanoid robots.
