Xpeng unveils ‘Physical AI’ products with robotaxi, flying cars

Xpeng held its 2025 AI Day event in Guangzhou, unveiling new products and technologies focused on “Physical AI” for future mobility.

The China-based EV maker introduced four applications—VLA 2.0, Robotaxi, Next-Gen IRON, and two flying systems from ARIDGE, each with planned mass production.

Xpeng also launched its VLA 2.0 large model, which the company said is designed for direct visual-to-action AI in vehicles, robots, and flying cars, with rollout to its Ultra models expected in Q1 2026.

Xpeng will open-source the VLA 2.0 model to business partners, naming Volkswagen as a launch customer.

The firm announced plans to release three Robotaxi models in 2026, saying these vehicles will feature Xpeng’s in-house developed AI stack and will not rely on lidar or high-definition maps.

A trial Robotaxi operation is also planned, though specific launch locations were not disclosed.

🔗 Source: Xpeng

🧠 Food for thought

Implications, context, and why it matters.

Xpeng’s 2026 camera-only robotaxis test China’s readiness

Xpeng says its 2026 robotaxis will run without lidar (light detection and ranging) or high-definition (HD) maps, unlike sensor fusion fleets from operators such as Waymo (Alphabet’s autonomous vehicle unit) 1, 2.
China restarted robotaxi permits in 2025 after a six-month pause tied to taxi-driver protests, signaling a 2035 autonomy leadership push 3.
Camera-only at scale remains unproven, with Tesla’s vision approach without lidar or HD maps still debated and lacking evidence of parity with sensor-fusion rivals 2.
Trial success hinges on Chinese approval for camera-only Level 4 (L4) use (no human safety driver), since most fleets still need detailed HD maps with frequent remapping 2.

Open VLA 2.0 aids third-party tools

Xpeng plans to open-source VLA 2.0 to business partners, and will offer an open SDK (software development kit) with mapping firm AutoNavi (a Chinese digital map provider) to support a wider self-driving community 4.
With a permissive license, the 720-billion-parameter VLA 2.0 could help outside teams build deployment tools plus software for safety testing and performance tuning 4.
Xpeng’s Turing AI chips deliver 2,250 TOPS (tera operations per second) in consumer vehicles and 3,000 TOPS in robotaxis, which could spur startups that tune hardware with software for edge AI inference (running models in the vehicle rather than the cloud) 1.
Volkswagen is the first commercial customer, which could bring B2B (business-to-business) software licensing revenue, with projections of RMB 2-3 billion a year by 2027 if adoption expands beyond the automaker 5.