Story
XPENG Iron is the kind of humanoid program HumanoidRoster should track early: it comes from a real industrial company with manufacturing experience, autonomy infrastructure, and a reason to care about embodied AI beyond a fundraising deck.
That does not make Iron proven. The public story is still mostly XPENG-controlled launch material and Physical AI positioning. XPENG now says it is aiming for year-end mass production, Q1 2027 showroom-guide deployment in XPENG retail outlets, and commercial deliveries next year, and its CVPR 2026 news item adds more specific research-ecosystem visibility around X-World, X-Foresight, X-Cache, forthcoming X-Mind, and Physical-World Foundation Model work. The Foundation Model Team's TuringViT paper adds a concrete visual-foundation component for IRON's perception stack, but it does not benchmark the robot's end-to-end task performance. June 2026 CnEVPost and KrASIA personnel coverage about a robotics product/product-lead departure is worth tracking as execution-risk context around that schedule, but it is not a robot capability datapoint. There is not yet enough independent evidence to score it like a factory pilot, and the profile should avoid treating mass-production ambition, model-stack details, conference visibility, or personnel-change coverage as deployment reality.
The reason to include Iron now is signal. If automakers become serious humanoid builders, XPENG is one of the programs that could matter. The profile should stay conservative until public specs, customer use, or repeatable work-cell evidence catches up.
Reality check
IRON is notable because XPENG is a serious EV and autonomy company bringing its Physical AI stack into humanoids. The score stays conservative because the public evidence is still launch/demo material, not independently verified production use or customer deployment.
- XPENG has an official IRON humanoid robot program page
- XPENG frames IRON as part of its broader Physical AI / mobility technology stack
- XPENG's AI Day article says Next-Gen IRON has a humanoid spine, bionic muscles, flexible skin, and 82 degrees of freedom
- IRON is presented publicly by a large, publicly traded EV maker rather than an anonymous concept team
- XPENG featured IRON in March 2026 public ecosystem events alongside its broader AI mobility stack
- XPENG's Q1 2026 results release said the company strives to mass-produce IRON by year-end, plans initial showroom deployment, and targets commercial deliveries next year
- XPENG's specific CVPR 2026 article says Physical-World Foundation Model work includes X-World, X-Foresight, X-Cache, and forthcoming X-Mind, and frames VLA 2.0/world-model work as relevant to VLA 2.0, Robotaxi, and humanoid robots
- XPENG says IRON is in hardware/software integration with mass production targeted by end-2026 and initial deployment as in-store shopping guides in XPENG retail outlets starting Q1 2027
- XPENG's Foundation Model Team says TuringViT has been integrated into the foundation model of the XPENG IRON robot as a visual foundation for object categories, spatial relations, manipulation-relevant regions, occlusions, dynamic scene changes, language alignment, and action planning
- Factory usefulness and mass-manufacturing readiness are not yet independently proven
- Public material does not yet provide a stable, complete spec sheet suitable for high-confidence scoring
- XPENG's May 2026 mass-production and commercial-delivery language is a forward-looking company plan, not evidence of completed manufacturing, delivered units, customer operation, task performance, or autonomy
- CnEVPost and KrASIA reported a robotics product/product-lead departure while XPENG was publicly targeting 2026 IRON mass production; treat this as execution-risk context only, not a robot capability datapoint
- XPENG's CVPR/world-model article and Q1 2027 showroom-guide roadmap are official forward-looking research and commercialization context, not proof that IRON has been mass-produced, delivered, deployed, or independently benchmarked
- XPENG vehicle-stack VLA 2.0, Robotaxi, and assisted-driving metrics should not be transferred directly to humanoid autonomy performance
- TuringViT's integration is a company-authored technical claim; the paper benchmarks the visual model rather than IRON's end-to-end task success, deployment reliability, or autonomy, and should not be presented as demonstrated manipulation capability