XPENG released a video on 22 September showing its IRON humanoid handling three customers in a car showroom. The robot localizes whoever is speaking, switches between Mandarin and English, and remembers who was interested in what. The clip carries an R&D test version label.
Key takeaways
- A nine-microphone array localizes sound, lip-movement recognition identifies the speaker
- Three onboard Turing AI chips deliver up to 2,250 TOPS
- Customer profiles combine facial recognition, voiceprint features and conversation history
- The robot switches Mandarin and English with no mode-change command
- Mass production in Q4 2026, commercial deployments from 2027
How the robot finds the speaker
IRON carries a nine-microphone array on board. It localizes the sound source, while lip-movement recognition identifies which person present is actually talking.
The response splits across three movement scales: the head alone, the waist, or the legs. Three onboard Turing AI chips do all the computing.
Memory instead of a script
The memory layer is the more interesting part. A customer profile is built from facial recognition, voiceprint features and conversation history combined.
In the video IRON corrected a participant who used another tester's nickname. It recalled interest in specific car models even after the testers swapped seats and changed clothes.
Mandarin-English switching happens without any mode-change command. XPENG credits a multilingual foundation model, conversational post-training and balanced multilingual training data.
What the video does not show
The "R&D Test Version Demonstration" label matters here. These are selected takes, not operational data.
| Layer | Shown in the clip | Left open |
|---|---|---|
| Speaker localization | nine microphones and lip movement | accuracy with a larger group |
| Customer memory | recognition after seat and clothing changes | retention over weeks |
| Languages | Mandarin and English with no command | every other language |
| Identity | corrected a wrong nickname | procedure on mistaken identity |
| Performance | no data | conversational latency |
Why it matters
A profile built from biometrics needs consent, a retention policy and a procedure for getting it wrong, and the demo shows none of those layers.
XPENG is selling a capability here, not a deployment. A showroom is also the easiest possible environment: enclosed, quiet, a handful of people at a time. The gap will show at the first site with real footfall.
What's next
- IRON mass production is set for Q4 2026, with first deliveries in the same period
- Commercial deployments in XPENG stores and campuses from 2027 — the first data from outside a demo
- The production line started on 8 September and, as CnEVPost reports, IRON walks off it under its own power
