XPeng unveils a next-generation IRON humanoid with 82 degrees of freedom and sets end-2026 for mass production

XPeng, the Guangzhou electric-vehicle maker, presented a second-generation version of its IRON humanoid robot at its 2025 AI Day on 5 November, according to a company release. The event, held at the XPENG Science Park in Guangzhou, also covered a new driving model, a robotaxi programme and flying cars, but the humanoid carried the most aggressive timetable: XPeng said it aims to reach large-scale mass production of IRON by the end of 2026.

The company put the new robot at 82 degrees of freedom across the body. Each hand has 22 degrees of freedom and is the size of a human hand, which XPeng said it achieved with what it called the smallest harmonic joint in the industry. Onboard computing comes from three of XPeng's own Turing AI chips, rated by the company at 3,000 TOPS of effective computing power. XPeng also said IRON is the first robot to use an all-solid-state battery, a claim it did not support with capacity, runtime or supplier details.

On software, XPeng said IRON carries its first-generation physical world large model and combines three model types it labels VLT, VLA and VLM, covering vision-language-task planning, vision-language-action control and vision-language understanding. The company will open a software development kit to developers worldwide. It listed guided tours, shopping guidance and directing customer traffic as the first commercial uses, and named Baosteel as an ecosystem partner for inspection and other industrial applications. XPeng gave no price, no unit target and no height or weight.

The gap from the first IRON is wide. When XPeng introduced the robot at its AI Day on 6 November 2024, it described more than 60 joints and 200 degrees of freedom, a five-year development effort and internal use in its own factories and stores. The company did not explain whether the 2024 and 2025 degree-of-freedom counts use the same method, so the lower 2025 figure should not be read as a reduction in capability.

The robot shares its hardware base with XPeng's cars. The same release said its VLA 2.0 driving model was trained on nearly 100 million video clips, which the company equates to 65,000 years of human driving, using a 30,000-card computing cluster and a 72-billion-parameter base model that is retrained every five days. Volkswagen is the launch customer for VLA 2.0 and has been designated for the Turing chip, XPeng said. Its robotaxi, due in three models in 2026, uses four Turing chips for 3,000 TOPS, with Amap as the first ride-hailing partner.

XPeng can fund the effort from a large balance sheet. Its third-quarter results, published on 17 November, showed revenue of RMB20.38 billion, up 101.8% from a year earlier, deliveries of 116,007 vehicles, a 20.1% gross margin and a net loss of RMB0.38 billion. Cash stood at RMB48.33 billion at 30 September, and research and development spending rose 48.7% to RMB2.43 billion in the quarter. Chief executive He Xiaopeng used the event to recast XPeng as a physical AI and embodied intelligence company rather than a carmaker.

The announcement lands in a crowded field. In China, UBTech raised gross proceeds of about HK$3,109.04 million on 25 November by placing 31,468,000 new H shares at HK$98.80, and said about 75% would go to investments or acquisitions along its value chain, according to its Hong Kong exchange filing. Abroad, Figure AI closed a US$1 billion Series C at a US$39 billion post-money valuation in September, while Apptronik, Agility Robotics and 1X Technologies are funded by venture rounds of US$350 million, US$150 million and US$100 million respectively.

In plain terms, XPeng is betting that the parts it already makes at scale for cars, namely inference chips, perception models trained on fleet video and battery packs, transfer to a walking machine. Degrees of freedom count independently controlled joints; more of them, especially in the hands, allow finer grasping but add actuators, cost and failure points. TOPS measures peak inference throughput, and XPeng's use of the word effective, without a published definition, makes direct comparison with rival platforms difficult.

The main risks are commercial. Shopping guidance and guided tours are low-value tasks that wheeled or tabletop robots already handle, so IRON's early uses do not yet prove the economics of a full humanoid. The solid-state battery claim is unverified, and no external customer beyond Baosteel's partnership was named.

The next milestones are dated by XPeng itself. VLA 2.0 reaches pioneer users in December 2025 and all Ultra models in the first quarter of 2026, the robotaxi models arrive in 2026, and IRON's mass-production target falls at the end of 2026.