Skin and body perception
It has a skin-like surface and AI to understand people and space.
Model not named
XPENG
How XPENG IRON works
XPENG IRON runs VLM and VLT on top of its own sensing and whole-body control — the parts the 4 public sources on this page actually name.
IRON is described with three named model layers: one that understands what it sees and hears, one that decides the task steps, and one that predicts the physical world and turns that into actions.
What XPENG IRON uses to sense, think, move and learn
Senses
Turns cameras, audio and touch into one picture
- VLM
Understands
Works out what the job is
- VLM
- VLT: Vision-Language-Task
Predicts
Anticipates what happens next
- XPENG VLA 2.0
- X-World
- X-Foresight
- X-Mind
Acts
Chooses the movement and the grasp
- XPENG VLA 2.0
- VLA 2.0 cross-domain model
Controls
Keeps the body balanced while it works
- Model not named
Learns
Improves the models between runs
- X-World
- X-Foresight
- X-Mind
What it learns feeds back into the models before the next run.
Greyed cards are parts the vendor hasn't named a model for.
01
How the robot is built to work
XPENG also has separate car-side world-model work that is not confirmed to run on IRON itself.
Technical read
XPENG states IRON combines VLM (vision-language understanding), VLT (Vision-Language-Task reasoning and decision-making) and VLA 2.0, which XPENG describes as both an action-generative model and a physical-world model spanning cars, humanoids and flying vehicles.
- Named VLM/VLT/VLA 2.0 stack
- Cross-domain physical-world model
- High DoF body
- 12
- System parts on the public record
- 09
- Parts with a public model name
- 04
- Primary architecture sources
4 confirmed as running on the robot
0 supplied by a partner
13 open questions left unanswered
XPENG is architecturally interesting because it discloses VLM/VLT/VLA 2.0 around IRON and separately discloses world-model work. Keep X-World/X-Mind as XPENG physical-AI context unless directly tied to IRON runtime.
02
How it senses, reasons and moves
Four-level chain
Understanding, task reasoning, prediction and action are named as separate models handing off in sequence.
Run the models
IRON likely has strong onboard compute, but exact TOPS should be quoted carefully.
- Turing AI chips
Contact safety
Skin could help contact awareness, but hard safety proof is not public.
Model not named
See and understand language
This is the part that understands what it sees and what people say.
- VLM
Decide the task steps
VLT is the robot task brain: it decides what job steps make sense.
- VLT: Vision-Language-Task
Predict the world and choose movement
XPENG is trying to make the AI predict the physical world before acting, and this same model turns that understanding into actions.
- XPENG VLA 2.0
Move around
The same physical-world AI thinking from cars may help the robot move around.
- VLA 2.0 cross-domain model
Hands and manipulation
The body and hands are highly articulated and skin-covered for human-like interaction.
Model not named
Whole-body walking
The body is designed to move more like a person, including spine and foot motion.
Model not named
Physical body
We know the body is complex; motor details are less public.
Not disclosed
Feeds back into the models
Around the stack
Named on the public record, but not part of the runtime chain above.
Car-side world-model context
XPENG may reuse prediction ideas across cars and robots, but we should not claim IRON definitely runs X-Mind or X-World.
- X-World
- X-Foresight
- X-Mind
Cross-domain learning
XPENG can reuse data and model ideas from cars and robotics to improve physical AI, but this is platform-level evidence, not IRON-specific proof.
- X-Mind
03
What sensors and hardware it has
Parts sit on the body only where the public record places them. 1 of 8 entries are still not publicly disclosed.
Vision and audio
01Electronic skin; exact camera/tactile sensor array not fully listed in text.
Skin/contact is public; full sensor list is not.
Reported for this robot · 1 source
Touch and force
01Seamless electronic skin and full-body soft exterior.
Skin-like covering may help safety/contact and human-facing design.
Confirmed on this robot · Across the robot · 1 source
Body position and balance
02Official XPENG IRON page lists 173 cm and 65 kg.
Human-sized but relatively light for its DoF.
Confirmed on this robot · Across the robot · 1 source
Still open: Some secondary pages list different early prototype specs; use official page for current IRON.
Bionic lumbar spine and bionic forefoot.
Spine and feet are designed to support more natural motion.
Confirmed on this robot · Legs · 1 source
Joints and actuators
0182 DoF full body; 22 DoF per hand.
Very high articulation, especially in the hands.
Confirmed on this robot · Across the robot · 1 source
Compute and connectivity
02VLM + VLT + VLA 2.0 described for IRON.
It has named model layers from seeing/talking to task/action.
Confirmed on this robot · 1 source
XPENG Turing AI compute is part of the broader physical-AI story, but exact IRON compute values need careful versioning.
Don't overquote TOPS unless the specific IRON page states it.
Conflicting public disclosures · 1 source
Still open: Need exact product spec.
Battery and runtime
01Not clearly specified in reviewed official IRON page text.
We do not know runtime.
Not publicly disclosed · 1 source
04
How it does one real job
Vendor-stated scenario
Retail/service task: guide a visitor to a display and hand over a small product sample
This walkthrough is a reasoned synthesis of publicly disclosed architecture pieces, not a confirmed end-to-end demo transcript.
01
Understand visitor request
The robot interprets what the visitor sees and says.
Technical detail
VLM processes visual/language context from the visitor.
- Senses
- Understands
- Predicts
- Acts
- Controls
- Learns
- VLM
Confirmed on this robot · 1 source
02
Choose task sequence
It converts the request into an ordered task plan.
Technical detail
VLT converts the request into a robot task plan: guide, stop, pick product, hand over.
- Senses
- Understands
- Predicts
- Acts
- Controls
- Learns
- VLT: Vision-Language-Task
Confirmed on this robot · 1 source
03
Predict and select action
It predicts likely outcomes and picks actions accordingly.
Technical detail
VLA 2.0 generates actions and includes physical-world understanding/prediction.
- Senses
- Understands
- Predicts
- Acts
- Controls
- Learns
- XPENG VLA 2.0
Confirmed on this robot · 1 source
04
Walk to display
The robot walks across the space toward the display.
Technical detail
Locomotion stack moves IRON through the space.
Autonomous navigation evidence should be tracked.
- Senses
- Understands
- Predicts
- Acts
- Controls
- Learns
- VLA 2.0 cross-domain model
Confirmed on this robot · 1 source
05
Pick sample
The hand grasps the item using vision, action and touch.
Technical detail
Hand/action system grasps item; electronic skin/contact may support safe interaction.
Grip/tactile resolution not public.
- Senses
- Understands
- Predicts
- Acts
- Controls
- Learns
No named model for this moment.
Confirmed on this robot · 2 sources
06
Hand over safely
The robot coordinates approach and handover while watching for contact.
Technical detail
Task/action layers coordinate approach and handover, while the contact/safety layer should detect interaction.
Safety details not public.
- Senses
- Understands
- Predicts
- Acts
- Controls
- Learns
- VLT: Vision-Language-Task
- XPENG VLA 2.0
Cyborgs interpretation · 2 sources
Parts are lit where the vendor's own description puts them to work. Unlit parts still run on the robot; they are simply not what this job turns on.
05
How it learns and improves
01
The robot works
Runs the parts of the system that later receive updates (1)
02
Experience is captured
Vehicle sensor/simulation data
03
Car-side world-model context
Vendor-stated, pre-production
04
Updates go back on the robot
Predicted future video/action, vehicle-side
Station 04 returns to station 01 — the loop repeats.
XPENG may reuse prediction ideas across cars and robots, but we should not claim IRON definitely runs X-Mind or X-World.
Technical detail
X-World is a controllable generative world model for future video/action simulation; X-Mind/X-Foresight are predictive world-model reasoning for vehicle-side driving. Strongest public evidence is in the vehicle/autonomy context, not confirmed as IRON runtime.
Vendor-stated, pre-production
- X-World
- X-Foresight
- X-Mind
Does IRON directly use X-Mind/X-World?
What the updates land on
- Predict the world and choose movementPredicts
06
What the vendor has shared
- 11
- Vendor-confirmed
- 03
- Reported or contested
- 04
- Stated, not shown
- 02
- The record is silent
A map of the public record, not a verdict. A silent line means the vendor has not said — it never moves the ranking.
What each band means
Vendor-confirmed
The vendor publicly describes this for the exact robot named on this page.
Reported or contested
Public sources disagree; we show the conflict rather than pick one.
Stated, not shown
The vendor has stated the intent; it has not been shown in a shipped configuration.
The record is silent
The public record does not say. We leave it visible as unknown.
Still open · 03
Does IRON directly use X-Mind/X-World?
XPENG has world-model work; direct IRON runtime linkage is not fully established.
Why it matters
Avoids confusing vehicle world model with humanoid runtime.
What is the exact compute/battery?
Current official IRON page has body/DoF details but limited compute/runtime disclosure.
Why it matters
Affects autonomy and deployment practicality.
What tasks has IRON done autonomously?
Public claims and demos need evidence classification separately from architecture disclosure.
Why it matters
Architecture is impressive but ranking needs proof.
07
Technical details
The record
Every row is the public position. Undisclosed rows stay in the list.
08
Sources
4 primary sources sit behind this page. Where they stop, the page says so.
Vendor
The ranking measures demonstrated capability. This page explains the public system behind it. The system description does not affect the score.
Return to the XPENG IRON report →