Xpeng advances AI for robotaxis, humanoid robots

Xpeng, a Chinese EV maker, plans to launch street trials of robotaxis and aims to start mass-producing humanoid robots in the second half of 2026, founder He Xiaopeng said at an event in Guangzhou.

The company is focusing on integrating in-house AI technology, including its “Turing” chip, into vehicles and robotics to compete in China’s crowded auto market.

Xpeng also unveiled four updated car models featuring new software, such as 3D navigation, and advanced hazard alerts.

Several Chinese automakers, including Li Auto, have increased investments in AI, as competition and a price war continue to affect the country’s auto sector.

Xpeng reported a net loss of 380 million yuan (US$54.42 million) in Q3 2025, with its CEO aiming for the company to break even by the end of 2025.

🔗 Source: Reuters

🧠 Food for thought

Implications, context, and why it matters.

Xpeng’s robotaxi timeline hinges on unclear municipal permits

Xpeng got an L3 road test license in Guangzhou on December 16, 2025, which allows conditional automated driving tests on highways and expressways 1. The piece does not state whether it has city permits for commercial robotaxis or street trials, and cities in China issue separate approvals for robotaxi pilots.
After the Ministry of Industry and Information Technology (MIIT) issued first L3 approvals for models from Changan Automobile plus Arcfox, Xpeng secured L3 road test permits in Guangzhou 2. It plans to launch three robotaxi models in 2026 with pilots focused on vehicle and tech development, not fleet operations 1.

AI toolchain vendors can target Xpeng’s Turing chip lineup

Xpeng plans to ship three Turing chips totaling 2,250 Tera Operations Per Second (TOPS) starting in Q3 2025 across higher-end Ultra variants 1. That gives independent AI vendors a chance to integrate, since the chip can run models with up to 30 billion parameters 3.
The Turing chip is Xpeng’s in-house automotive AI processor and integrates two proprietary Neural Processing Units (NPUs) with a 40-core processor 4. Xpeng claims 20% higher utilization than general-purpose automotive chips 3, which could leave room for third-party vendors that focus on model compression and efficient deployment.
Xpeng’s Vision-Language-Action (VLA) 2.0 removes intermediate language translation and maps visual inputs to driving actions 1. Vendors that speed up end-to-end inference on Turing could land partnerships, as Xpeng projects automakers may spend 50 billion yuan on Research and Development, with 30 billion yuan for AI 1.