How they work

How humanoid robots work

Every robot has to do the same six things. These pages show which AI models each vendor has actually named for each one.

12 robots · 40 named models

Senses

Turns cameras, audio and touch into one picture

Understands

Works out what the job is

Predicts

Anticipates what happens next

Acts

Chooses the movement and the grasp

Controls

Keeps the body balanced while it works

Learns

Improves the models between runs

What it learns feeds back into the models before the next run.

Greyed cards are parts the vendor hasn't named a model for.

Every robot we have written up

The robots whose vendors name the most models come first.

1X Technologies
1X NEO
NEO understands you through a built-in language assistant, uses a dedicated robot-body model called Redwood to move and do chores, and separately researches a 'world model' that can imagine and evaluate outcomes. When it does not know how to do something, a remote human expert can step in.
1X NEO system loop: Senses — Model not named; Understands — Model publicly named; Predicts — Model publicly named; Acts — Model publicly named; Controls — Model publicly named; Learns — Model publicly named.
05 of 06 named
Read how it works →
AgiBot
AgiBot A2
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. 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.
AgiBot A2 system loop: Senses — Model publicly named; Understands — Model publicly named; Predicts — Model publicly named; Acts — Model publicly named; Controls — Model not named; Learns — Model publicly named.
05 of 06 named
Read how it works →
XPENG
XPENG IRON
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. XPENG also has separate car-side world-model work that is not confirmed to run on IRON itself.
XPENG IRON system loop: Senses — Model publicly named; Understands — Model publicly named; Predicts — Model publicly named; Acts — Model publicly named; Controls — Model not named; Learns — Model publicly named.
05 of 06 named
Read how it works →
Apptronik
Apptronik Apollo
Apollo 2 is Apptronik's own humanoid body and control stack, called Artemis, which handles perception, planning, safety and human interaction. Apptronik also has a training-data partnership with Google DeepMind's Gemini Robotics models, but the product page centers Artemis, not Gemini, as what actually runs the robot.
Apptronik Apollo system loop: Senses — Model publicly named; Understands — Model publicly named; Predicts — Not disclosed; Acts — Model publicly named; Controls — Model publicly named; Learns — Partner technology.
04 of 06 named · 01 partner · 01 undisclosed
Read how it works →
Figure AI
Figure 03
Figure uses a three-layer system: one layer understands the job, one turns everything the robot sees and feels into movement, and one handles fast balance and actuator control.
Figure 03 system loop: Senses — Model not named; Understands — Model publicly named; Predicts — Not disclosed; Acts — Model publicly named; Controls — Model publicly named; Learns — Model not named.
03 of 06 named · 01 undisclosed
Read how it works →
UBTECH
UBTECH Walker S
Walker S is an industrial humanoid built for factory work. It understands spoken instructions with an LLM, builds a 3D map of the factory floor to walk around safely, and uses depth cameras plus force-feedback joints to find and grasp objects. UBTECH's newer Walker S1 and S2 pages describe extra planning and learning technology, but that is family-level evidence, not confirmed proof about the exact Walker S.
UBTECH Walker S system loop: Senses — Model not named; Understands — Model publicly named; Predicts — Not disclosed; Acts — Model not named; Controls — Model not named; Learns — Model publicly named.
02 of 06 named · 01 undisclosed
Read how it works →
Fourier Intelligence
Fourier GR-2
GR-2 is a highly articulated humanoid body that developers can control directly, or drive by hand using a VR/teleoperation system to record examples for training robot-learning models later. Fourier does not publicly name an onboard model that runs the robot autonomously.
Fourier GR-2 system loop: Senses — Model not named; Understands — Not disclosed; Predicts — Not disclosed; Acts — Model publicly named; Controls — Model not named; Learns — Model not named.
01 of 06 named · 02 undisclosed
Read how it works →
Boston Dynamics
Atlas (electric)
Atlas is Boston Dynamics' fully electric, highly dynamic industrial humanoid, built for factory material handling. It can be run autonomously, teleoperated, or steered from a tablet. Boston Dynamics has not publicly named a language model, task planner, action model or world model behind it — the public record centers on the body's dynamics and control.
Atlas (electric) system loop: Senses — Model not named; Understands — Model not named; Predicts — Not disclosed; Acts — Model not named; Controls — Model not named; Learns — Not disclosed.
No models publicly named · 02 undisclosed
Read how it works →
Kepler Robotics
Kepler Forerunner K2
Kepler shows a strong-looking industrial humanoid body — height, weight and actuator strength are public — but Kepler has not publicly named the software that understands instructions, plans tasks, predicts outcomes or decides what the hands should do.
Kepler Forerunner K2 system loop: Senses — Model not named; Understands — Model not named; Predicts — Not disclosed; Acts — Model not named; Controls — Model not named; Learns — Not disclosed.
No models publicly named · 02 undisclosed
Read how it works →
NEURA Robotics
NEURA 4NE-1
4NE1 is pitched by NEURA as a 'cognitive' humanoid that can understand language, see its environment and learn actions through reinforcement learning, moving through unstructured factory or service spaces with exchangeable arms. NEURA has not named the exact vision-language-action model or world model behind this behaviour.
NEURA 4NE-1 system loop: Senses — Model not named; Understands — Model not named; Predicts — Not disclosed; Acts — Model not named; Controls — Model not named; Learns — Model not named.
No models publicly named · 01 undisclosed
Read how it works →
Tesla
Tesla Optimus
Tesla is openly building a full stack for Optimus — senses, a task brain, a movement brain and a balance system — but it has not put public names on most of these models, unlike some competitors.
Tesla Optimus system loop: Senses — Model not named; Understands — Model not named; Predicts — Not disclosed; Acts — Model not named; Controls — Model not named; Learns — Model not named.
No models publicly named · 01 undisclosed
Read how it works →
Unitree
Unitree H2
H2 is a tall, powerfully actuated humanoid body with a very capable onboard computer, sold as a platform that can run 'diverse intelligent models'. Unitree has not publicly named the reasoning, perception or action models that would make it act autonomously.
Unitree H2 system loop: Senses — Not disclosed; Understands — Model not named; Predicts — Not disclosed; Acts — Model not named; Controls — Model not named; Learns — Model not named.
No models publicly named · 02 undisclosed
Read how it works →

How the system works, and what it has proved

These pages describe the system. The monthly ranking measures what has been demonstrated.

See the monthly ranking All humanoid reports How we score