AgiBot / Zhiyuan Robotics
See and understand indoor spaces
A2 has sensors that help it see and understand indoor spaces and objects.
- A2 perception system
AgiBot
How AgiBot A2 works
AgiBot A2 runs ActionGPT and A2 perception system on top of its own sensing and whole-body control — the parts the 4 public sources on this page actually name.
A2 is a full-size service/logistics humanoid built around AgiBot's own ActionGPT movement-planning model and HIMUS indoor navigation, with enterprise knowledge features for business use.
What AgiBot A2 uses to sense, think, move and learn
Senses
Turns cameras, audio and touch into one picture
- A2 perception system
Understands
Works out what the job is
- ActionGPT
- GO-2 platform model
Predicts
Anticipates what happens next
- Genie Envisioner 2.0
- GE-Sim
Acts
Chooses the movement and the grasp
- HIMUS autonomous navigation
- GO-2 platform model
Controls
Keeps the body balanced while it works
- Model not named
Learns
Improves the models between runs
- Genie Envisioner 2.0
- GE-Sim
- AgiBot World dataset
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
AgiBot also has separate platform-level work — the GO-2 embodied-AI model and the Genie Envisioner world-model/simulator — that is not confirmed to run inside A2 itself.
Technical read
AgiBot states A2 combines ActionGPT (task-to-movement generation), HIMUS autonomous navigation and an enterprise knowledge/RAG-style stack for service scenarios. The wider AgiBot platform includes GO-2, an embodied-AI foundation model unifying vision, language and action, and Genie Envisioner, an action-driven world-model/simulator platform; both are disclosed at the AgiBot platform level rather than confirmed as A2 runtime components.
- Named ActionGPT + HIMUS stack
- Enterprise knowledge/RAG
- Platform-level GO-2 / world model context
- 12
- System parts on the public record
- 09
- Parts with a public model name
- 04
- Primary architecture sources
2 confirmed as running on the robot
0 supplied by a partner
12 open questions left unanswered
A2 has direct product-page evidence for ActionGPT and HIMUS. GO-2 and Genie Envisioner are AgiBot platform models; link them as platform direction unless AgiBot explicitly says a given A2 unit runs them.
02
How it senses, reasons and moves
Two-level split
A thinking layer sets the intent; a second layer turns it into movement.
AgiBot / Zhiyuan Robotics
Use business knowledge
A2 can refer to business information rather than only generic conversation.
Model not named
AgiBot / Zhiyuan Robotics
Work safely around people
A2 is meant to work around people, but hard safety proof needs more documents.
Model not named
AgiBot / Zhiyuan Robotics
Talk and understand requests
A2 can talk with people and use business information to answer questions or start tasks.
Model not named
AgiBot / Zhiyuan Robotics
Turn requests into movement plans
ActionGPT is the layer that turns a request into a sensible sequence of robot movements.
- ActionGPT
AgiBot / Zhiyuan Robotics
Navigate the building
HIMUS is the navigation brain that helps A2 move through a building.
- HIMUS autonomous navigation
AgiBot / Zhiyuan Robotics
Reach, grasp and carry
A2 has enough controlled joints for full-body movement and arm/hand work.
Model not named
AgiBot / Zhiyuan Robotics
Walk and balance
The walking system handles balance and movement, including long-endurance walking in the A2 Ultra variant.
Model not named
AgiBot / Zhiyuan Robotics
Physical joints
The physical muscles are modular joints, but exact internals are not fully public.
Model not named
Feeds back into the models
Around the stack
Named on the public record, but not part of the runtime chain above.
AgiBot / Zhiyuan Robotics
Practice in a modelled world (platform-level)
AgiBot is building a way for robots to practice or predict actions inside a modelled world, but this is platform-level work, not confirmed to run inside A2 itself.
- Genie Envisioner 2.0
- GE-Sim
AgiBot / Zhiyuan Robotics
AgiBot's broader vision-language-action brain (platform-level)
AgiBot has a broader robot-action brain, GO-2, that connects seeing, language and movement across its platform — but A2 itself is specifically named as using ActionGPT, not GO-2.
- GO-2 platform model
AgiBot / Zhiyuan Robotics
Improve future robots with data (platform-level)
AgiBot is building a data engine — the AgiBot World datasets plus GO/GE platform models — so future robots can learn from many tasks, but this is platform-level evidence, not proof about a specific A2 unit.
- AgiBot World dataset
03
What sensors and hardware it has
Parts sit on the body only where the public record places them. 3 of 8 entries are still not publicly disclosed.
Vision and audio
01Product page describes embodied/visual capabilities, but exact camera/tactile/force sensor list is not complete in text.
We know it sees and navigates, but not every sensor SKU.
Not publicly disclosed · 1 source
Still open: Need an official sensor table.
Body position and balance
03169 cm height listed on the official A2 page.
Close to human height, suited to indoor spaces.
Confirmed on this robot · Across the robot · 1 source
69 kg weight listed on the official A2 page.
Heavy enough for stability but still human-scale.
Confirmed on this robot · Across the robot · 1 source
60 cm turning radius listed on the A2 page; HIMUS navigation named.
It can turn in tight indoor spaces.
Confirmed on this robot · Legs · 1 source
Hands and end effectors
01Exact hand DoF and payload are variant-specific and incompletely disclosed in public text.
We should avoid inventing finger details for A2.
Not publicly disclosed · Hands · 1 source
Still open: Confirm differences between A2, A2-W and A2 Ultra.
Joints and actuators
0140+ active degrees of freedom listed on the official A2 page.
Many controllable body joints for varied movement.
Confirmed on this robot · Across the robot · 1 source
Still open: Variant differences are possible.
Compute and connectivity
01Some secondary sources mention high compute, but the reviewed official A2 text did not expose the chip or TOPS figures clearly.
Do not assume the chip until official specs are found.
Not publicly disclosed · Torso · 1 source
Still open: Verify with an official spec sheet or technical brochure.
Battery and runtime
01700 Wh battery, 2 hour runtime, with battery-swapping support, per the official A2 page.
It can work for a limited shift and swap batteries to keep going.
Confirmed on this robot · 1 source
Still open: Runtime will vary by task.
04
How it does one real job
Vendor-stated scenario
Service/logistics task: greet a visitor, identify a parcel, and deliver it to a marked indoor station
This walkthrough is a reasoned synthesis of publicly disclosed architecture pieces, not a confirmed end-to-end demo transcript.
01
Understand request
A person asks for a parcel delivery; the robot understands the request and any enterprise metadata.
Technical detail
Language interface plus enterprise knowledge stack interpret the request and link it to business metadata.
Exact speech/LLM details are unclear.
- Senses
- Understands
- Predicts
- Acts
- Controls
- Learns
No named model for this moment.
Confirmed on this robot · 1 source
02
Break into subtasks
ActionGPT converts the request into movement/action steps: find parcel, navigate, grasp, deliver.
Technical detail
ActionGPT generates a task-to-movement sequence from the interpreted request.
- Senses
- Understands
- Predicts
- Acts
- Controls
- Learns
- ActionGPT
Confirmed on this robot · 1 source
03
Navigate to parcel
The navigation stack plans a route and moves through the indoor space.
Technical detail
HIMUS navigation plans a route; locomotion executes walking and balance toward the target.
HIMUS internal details are high-level.
- Senses
- Understands
- Predicts
- Acts
- Controls
- Learns
- HIMUS autonomous navigation
Confirmed on this robot · 1 source
04
Identify parcel
Vision recognizes the parcel or marker; the exact perception model is not public.
Technical detail
Perception system processes visual input to identify the target object or marker.
Identification mechanism needs further evidence.
- Senses
- Understands
- Predicts
- Acts
- Controls
- Learns
- A2 perception system
Cyborgs interpretation · 1 source
05
Pick and carry
Arm/hand control grasps the parcel while locomotion maintains stability.
Technical detail
Manipulation system grasps the parcel using joint and visual-target signals while whole-body control preserves balance.
Need public videos with task duration for confirmation.
- Senses
- Understands
- Predicts
- Acts
- Controls
- Learns
No named model for this moment.
Cyborgs interpretation · 1 source
06
Deliver and interact
The robot reaches the station, announces or confirms delivery, and executes placement.
Technical detail
Navigation completes the route while language and action layers confirm delivery and execute placement.
- Senses
- Understands
- Predicts
- Acts
- Controls
- Learns
- ActionGPT
- HIMUS autonomous navigation
Cyborgs interpretation · 1 source
07
Improve over time
The fleet/data ecosystem can improve future models, but the exact A2 update loop is not public.
Technical detail
AgiBot World data and GO/GE platform models feed a broader learning flywheel that is not confirmed to directly update deployed A2 units.
Not direct task proof for a specific A2 unit.
- Senses
- Understands
- Predicts
- Acts
- Controls
- Learns
- AgiBot World dataset
- GO-2 platform model
- Genie Envisioner 2.0
- GE-Sim
Confirmed for the robot family · 3 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 (2)
02
Experience is captured
Vision/language/action data · Robot trajectories
03
Practice in a modelled world (platform-level)
Confirmed for the robot family
04
Updates go back on the robot
Generated futures/simulated training/evaluation
Station 04 returns to station 01 — the loop repeats.
AgiBot is building a way for robots to practice or predict actions inside a modelled world, but this is platform-level work, not confirmed to run inside A2 itself.
Technical detail
AgiBot describes Genie Envisioner as an action-driven world-model platform and simulator for learning within model-generated worlds, used for training/evaluation rather than disclosed as an onboard A2 runtime component.
Confirmed for the robot family
- Genie Envisioner 2.0
- GE-Sim
- AgiBot World dataset
Exact usage inside A2 runtime is not public.
What the updates land on
- AgiBot's broader vision-language-action brain (platform-level)Understands
- Practice in a modelled world (platform-level)Predicts
06
What the vendor has shared
- 16
- Vendor-confirmed
- 01
- Stated, not shown
- 03
- 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.
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
Which A2 variant is being ranked?
Screenshots reference AgiBot A2; the official page lists baseline A2 specs.
Why it matters
A2, A2 Lite, A2 Ultra, A2-W and A2 Max can differ materially in specs and capability.
Is GO-2 deployed on A2?
GO-2 is public at the platform level; exact A2 runtime linkage is not explicit.
Why it matters
GO-2 is important for reasoning/action, but platform evidence is not the same as confirmed product runtime.
What exact sensors/hands are on A2?
Official text does not expose a full sensor/hand bill of materials.
Why it matters
Tactile and manipulation capability depends on hardware detail that is not published.
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 AgiBot A2 report →