Unitree Robotics
See and sense the world
The page tells us H2 can run AI models, not exactly what sensors it uses to see and sense its surroundings.
Not disclosed
Unitree
How Unitree H2 works
Unitree H2's vendor has not publicly named the models behind it, so this page sets out what is disclosed: the sensing, control and hardware, with the evidence for each part.
H2 is a tall, powerfully actuated humanoid body with a very capable onboard computer, sold as a platform that can run 'diverse intelligent models'.
What Unitree H2 uses to sense, think, move and learn
Senses
Not disclosed
- Not disclosed
Understands
Works out what the job is
- Model not named
Predicts
Not disclosed
- Not disclosed
Acts
Chooses the movement and the grasp
- Model not named
Controls
Keeps the body balanced while it works
- Model not named
Learns
Improves the models between runs
- Model not named
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
Unitree has not publicly named the reasoning, perception or action models that would make it act autonomously.
Technical read
Unitree's public H2 page centres on motion-control algorithms, 31 DoF, 360 N.m joint torque and a 2070 TOPS chip described as supporting diverse embodied models, with algorithms evolving via OTA updates. No H2-specific VLA, embodied reasoner, world model or sensor-suite table is named in reviewed official material.
- Motion/hardware disclosed
- Compute platform for external models
- No named H2 intelligence stack
- 13
- System parts on the public record
- 02
- Parts with a public model name
- 02
- Primary architecture sources
1 confirmed as running on the robot
0 supplied by a partner
14 open questions left unanswered
H2 is a strong 'body and motion' case. The official page emphasises motion-control algorithms and high compute for embodied models. Do not assume every external model or H2 Plus/EDU configuration applies to base H2.
02
How it senses, reasons and moves
Stack not publicly named
The vendor describes what each part does, but has not named the models that run it.
Unitree Robotics
Run the models
H2 has a very powerful onboard AI computer, described as ready to run 'diverse intelligent models'.
Model not named
Unitree Robotics
Understand instructions
The robot can support AI models, but Unitree has not named a ChatGPT-like H2 brain that turns spoken or written instructions into tasks.
Not disclosed
Unitree Robotics
Stay within limits
We don't yet know the full safety envelope around people from official H2 material.
Not disclosed
Unitree Robotics
Predict what may happen
No public 'imagine possible futures' model is named for H2.
Not disclosed
Unitree Robotics
Decide what to do
H2 is described as a platform for models rather than a robot with a fully disclosed reasoning brain built in.
Not disclosed
Unitree Robotics
Find its way around
H2 moves very well, but the map/route-planning brain behind getting from place to place is not clearly named.
Not disclosed
Unitree Robotics
Choose an action
H2 has the computer to run robot AI 'brains', but the exact model choosing its actions is not publicly named.
Not disclosed
Unitree Robotics
Grip and handle objects
We know H2 is humanoid, but Unitree's base H2 page does not spell out hand dexterity or tactile-sensing detail.
Not disclosed
Unitree Robotics
Move the body
This is Unitree's strongest publicly shown box: a body with many joints, strong actuators and motion-control software that keeps it stable and lets it perform precise, dynamic movement.
Model not named
Unitree Robotics
Powerful joints
The robot has powerful joints built for dynamic, human-like motion.
Model not named
Feeds back into the models
Around the stack
Named on the public record, but not part of the runtime chain above.
Unitree Robotics
Improve over time
Unitree can update H2's movement software over time and positions it for a broader, global community to help build embodied models.
Model not named
Unitree Robotics
Other Unitree humanoids
Unitree sells other humanoids (like G1 and H1) and H2 variants (Plus/EDU) that may use different sensors, hands or software. None of that should be assumed to apply to the base H2 reviewed here.
Model not named
03
What sensors and hardware it has
Parts sit on the body only where the public record places them. 4 of 9 entries are still not publicly disclosed.
Vision and audio
01Reviewed official text does not provide a full sensor list.
The exact camera, lidar or tactile package is unknown.
Not publicly disclosed · 1 source
Still open: An official sensor table for H2 is needed.
Body position and balance
01180 cm listed on official H2 page.
It is a tall, human-scale robot.
Confirmed on this robot · Across the robot · 1 source
Hands and end effectors
01Not clearly specified in official H2 text reviewed.
Do not assume dexterous, tactile hands for base H2.
Not publicly disclosed · Hands · 1 source
Still open: Base H2 vs research/Plus hand options need confirmation.
Joints and actuators
0231 degrees of freedom.
31 main controllable joints allow complex, dynamic movement.
Confirmed on this robot · Across the robot · 2 sources
360 N.m joint torque.
Strong joints power fast, dynamic movement.
Confirmed on this robot · Across the robot · 2 sources
Still open: The exact joint this figure refers to is not specified in reviewed text.
Compute and connectivity
022070 TOPS chip.
A powerful onboard computer for running AI models.
Confirmed on this robot · Torso · 1 source
OTA algorithm evolution mentioned.
Movement software can improve after the robot ships.
Confirmed on this robot · 1 source
Battery and runtime
01Not disclosed in official H2 text reviewed.
We don't know practical runtime between charges.
Not publicly disclosed · 1 source
Safety hardware
01Not fully disclosed in official H2 text reviewed.
We don't yet know the full human-proximity safety envelope.
Not publicly disclosed · 1 source
04
How it does one real job
Illustrative walkthrough
Research/industrial demo task: walk to a workbench, pick up a light item, and hand it to an operator
This is a reasoned illustration built from the official H2 page's hardware and compute claims, not a documented end-to-end demo of this exact handover task.
01
Load a skill or model
The robot loads the brain for this job.
Technical detail
An embodied model or task script is loaded onto the 2070 TOPS compute for the handover scenario; no specific model name is public.
No specific model name is public for this step.
- Senses
- Understands
- Predicts
- Acts
- Controls
- Learns
Cyborgs interpretation · 1 source
02
Walk to the bench
It walks smoothly to the table.
Technical detail
H2's vendor-confirmed motion-control stack drives the 31-DoF, 360 N.m-torque body to the workbench.
Dynamic-motion proof is stronger public evidence than manipulation proof for H2.
- Senses
- Understands
- Predicts
- Acts
- Controls
- Learns
Confirmed on this robot · 2 sources
03
Identify the item
It sees the object on the bench.
Technical detail
Cameras or other sensors likely locate the item; the exact sensor suite and perception model are not public.
H2's sensor suite is not fully disclosed.
- Senses
- Understands
- Predicts
- Acts
- Controls
- Learns
Cyborgs interpretation · 1 source
04
Reach and grasp
It reaches out and grabs the item.
Technical detail
An arm/hand controller executes the grasp; exact dexterity and tactile-sensing hardware for base H2 are unclear.
Hand and tactile specs for base H2 are needed to confirm this step.
- Senses
- Understands
- Predicts
- Acts
- Controls
- Learns
Cyborgs interpretation · 1 source
05
Balance during the handover
It does not fall over while reaching out.
Technical detail
The vendor-confirmed whole-body controller maintains balance as the arm extends and the payload shifts.
- Senses
- Understands
- Predicts
- Acts
- Controls
- Learns
Confirmed on this robot · 1 source
06
Release to the operator
It hands over the item to a person.
Technical detail
The robot releases the item when the human is assessed as ready; exact human-detection and safety logic are not public.
Human-proximity safety behaviour for this moment is not publicly detailed.
- Senses
- Understands
- Predicts
- Acts
- Controls
- Learns
Cyborgs interpretation · 1 source
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
Robot logs/model updates
03
Improve over time
Confirmed on this robot
04
Updates go back on the robot
Improved movement/models
Station 04 returns to station 01 — the loop repeats.
Unitree can update H2's movement software over time and positions it for a broader, global community to help build embodied models.
Technical detail
The official page states motion-control algorithms evolve via over-the-air updates and positions H2 for global co-creation of embodied models, distinct from onboard autonomous learning.
Confirmed on this robot
Model-update governance and training-data sourcing are unknown.
What the updates land on
- Move the bodyControls
06
What the vendor has shared
- 10
- Vendor-confirmed
- 12
- 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.
The record is silent
The public record does not say. We leave it visible as unknown.
Still open · 04
What named intelligence stack, if any, runs on H2?
The official H2 page states it supports diverse intelligent models but does not name an H2-specific reasoner, world model, VLA or perception stack.
Why it matters
Users will want to compare H2's brain to named stacks from other vendors, but Unitree markets H2 primarily as compute-ready hardware for external models.
What is H2's full sensor and hand/tactile configuration?
Reviewed official text does not provide a full sensor list, and hand/tactile detail is not clearly specified for base H2.
Why it matters
Perception and manipulation claims cannot be assessed without knowing the actual sensor and hand hardware.
Is Unitree using a named model such as UnifoLM or a WVLA-style policy on H2?
No official source ties such a model directly to H2 in reviewed material.
Why it matters
This would materially change how H2's action-choosing layer should be described.
What is H2's battery life and human-proximity safety architecture?
Neither is disclosed in reviewed official H2 text.
Why it matters
Practical runtime and safety envelope both affect real-world deployability.
07
Technical details
The record
Every row is the public position. Undisclosed rows stay in the list.
08
Sources
2 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 Unitree H2 report →