# How a humanoid robot works: actuators, compute, sensors and the model

> Of 182 humanoids on record, 178 say what moves the joints and 43 say what computer is inside. The further a layer sits from the metal, the less the maker tells you about it.

Published: 2026-09-06
Section: Explainer
Author: Geert Jan van Vlastuin, Editor, humanoid.observer
Canonical: https://humanoid.observer/blog/how-a-humanoid-robot-works-actuators-compute-sensors-and-the-model/
Licence: editorial text all rights reserved. The underlying data is CC BY 4.0 at https://humanoid.observer/data/.
Cite as: humanoid.observer, https://humanoid.observer/blog/how-a-humanoid-robot-works-actuators-compute-sensors-and-the-model/, accessed <date>.

---
A humanoid is four layers. Actuators move the joints. Sensors see and feel. A computer runs the maths. A model decides what to do next. Our 182 records describe the first layer 178 times and the third layer 43 times. Makers are precise about motors and vague about brains, and the vagueness sits exactly where a human operator can hide.

## The four layers and how often anyone describes them

Field coverage across the 182 records, with the most common answer.

| Layer | Field | Records filled | Most common answer |
|---|---|---|---|
| Actuators | `actuation` | 178 of 182 | "electric" with nothing else (87) |
| Joints | `dof` | 83 | 28 joints (6 machines); median 32 |
| Compute | `compute` | 43 | a generic "onboard computer" (7); Nvidia Jetson (6) |
| Sensors | `sensors` | 64 | a camera (47); an IMU (23) |
| Model | `ai_stack` | 78 | reinforcement-learning locomotion (17) |
| Human in the loop | `teleoperation` | 148 (34 unknown) | partial (90) |

Only 18 of the 182 fill all five technical fields, and 56 fill none of the four beyond actuation. Of the 44 machines you can order, 25 publish a joint count, 9 name the computer and 10 list the sensors. You are asked to buy a body whose spec sheet stops at the motors.

## Actuators: electric has won and hydraulics are in the museum

The actuator is the motor plus gearing at each joint. It sets what the robot can lift, how fast it moves and what wears out first. Of the 178 records that describe it, 170 are electric only, 7 mention hydraulics and 1 ([NAVER AMBIDEX](/robots/naver-ambidex/)) is cable-driven.

The hydraulic seven are mostly history. The [hydraulic Atlas](/robots/boston-dynamics-atlas-hydraulic/) from [Boston Dynamics](/companies/boston-dynamics/) had 28 hydraulic joints and was retired on 16 April 2024 after 11 years, replaced by the [electric Atlas](/robots/boston-dynamics-atlas-electric/). Its predecessor [PETMAN](/robots/boston-dynamics-petman/) was hydraulic and tethered. [WABOT-1](/robots/wabot-1/) from 1973 had hydraulic legs. [Nadia](/robots/ihmc-nadia/) at IHMC has some hydraulic joints. [Phoenix](/robots/sanctuary-phoenix/) from [Sanctuary AI](/companies/sanctuary-ai/) puts miniature hydraulic hands on an electric body. The last two are [Clone Robotics](/companies/clone-robotics/)' [Protoclone](/robots/clone-protoclone/) and [Clone Alpha](/robots/clone-alpha/), artificial muscles driven by a water pump; the Alpha has no delivered unit and no public price.

Electric wins because it runs on a battery, does not leak and is cheap in volume. The word on its own tells you nothing: 87 records say it and stop. Only 5 give a torque figure: the [Unitree G1](/robots/unitree-g1/) at up to 120 Nm in the knee, the [Unitree H1](/robots/unitree-h1/) at up to 360 Nm, [Fourier GR-1](/robots/fourier-gr-1/) and [Xiaomi CyberOne](/robots/xiaomi-cyberone/) at 300 Nm, [AgiBot Raise A1](/robots/agibot-raise-a1/) at 350 Nm. Five name linear actuators, among them [Optimus Gen 1](/robots/tesla-optimus-gen-1/) and [Apollo](/robots/apptronik-apollo/) from [Apptronik](/companies/apptronik/). Four are tendon-driven: the three [NEO](/robots/1x-neo/) generations from [1X](/companies/1x/) and Roboy. Six say in-house or custom. The torque number is the one that predicts payload, and 173 records do not give it.

## Joints: 83 publish a count and the famous machines do not

Degrees of freedom is the number of independently moving joints. 83 records give a figure. The minimum is 12 ([Berkeley Humanoid](/robots/berkeley-humanoid/)), the median is 32, the mean is 35.7. The maximum is 200 for Protoclone, a musculoskeletal android with over 1,000 artificial muscles and not comparable; leave it out and the top is [Xpeng Iron](/robots/xpeng-iron/) at 82, then [HopeJR](/robots/hugging-face-hopejr/) at 66, an open-source kit on preorder at a $3,000 target price. 11 machines are under 20 joints, 24 are in the twenties, 18 in the thirties, 18 in the forties and 12 at 50 or more.

The count is silent on where the joints go: Iron puts 22 of its 82 in each hand. We covered that in [why hands are the hardest part](/blog/why-hands-are-the-hardest-part/). The count is also silent on the best-known machines. [Figure 02](/robots/figure-02/), [Figure 03](/robots/figure-03/), [Optimus Gen 2](/robots/tesla-optimus-gen-2/), the electric Atlas, NEO, [Digit](/robots/agility-digit-v4/) and Apollo have no total joint count in our record. Among the 44 you can order, 25 publish one, with a median of 24. A high count is a bill of materials, not a capability: HopeJR publishes more joints than the electric Atlas record shows, and it is a hobby kit.

## Compute: 43 name a computer, 6 say Jetson, none say Thor

The computer runs the balance loop and, on newer machines, the network that turns camera frames into joint targets. 43 records name it. [Nvidia](/companies/nvidia/) Jetson appears 6 times: the G1 EDU, the H1 as an option, the [Unitree R1](/robots/unitree-r1/) EDU, [Booster T1](/robots/booster-t1/), [EngineAI PM01](/robots/engineai-pm01/) and [Hexagon AEON](/robots/hexagon-aeon/). Four of those specify the Orin generation. Jetson Thor, which Nvidia's own record says went on general sale in August 2025 at $3,499 for the developer kit, appears in no robot record at all.

Three run in-house silicon: Optimus Gen 1 and Gen 2 on a computer derived from [Tesla](/companies/tesla/)'s FSD chip, and Iron on three Xpeng Turing chips at about 2,250 TOPS. Only Iron and the G1 give a TOPS figure. Five name an Intel x86 part, from an Atom in [DARwIn-OP](/robots/darwin-op/) to a Core i7 in the H1. Two use a Raspberry Pi, two an Arduino, two a cloud brain ([EMIEW3](/robots/hitachi-emiew/) and AMBIDEX over 5G). Seven say only that there is an onboard computer. Figure 03's record says Helix runs onboard and details are not published.

The compute line decides who can run their own code. A Jetson with an open SDK is a computer you can log into; the G1 base ships an 8-core CPU and the 100-TOPS Orin NX is EDU only, the ladder in [the G1 article](/blog/unitree-g1-cheap-one-cannot-run-your-code/). An in-house chip is a computer the maker programs and you do not. 35 of the 44 buyable machines do not name theirs, so ask.

## Sensors: cameras on 47, touch on 8

Sensors are what the model gets to know. A camera gives pictures, a depth camera distance, lidar a 3D map for walking, an IMU the sense of falling, tactile sensors a feeling of grip. 64 records list sensors. 47 mention a camera, 23 an IMU, 17 a depth or stereo camera, 12 lidar, 15 microphones. 19 mention touch, force or pressure sensing, and 8 use the word tactile: Figure 03 with fingertip tactile sensors and palm cameras, Optimus Gen 2 with tactile fingertips and foot force sensors, [Fourier GR-3](/robots/fourier-gr-3/) with 31 pressure sensors in a soft skin. Five list cameras and nothing else.

The mix tells you what the maker thinks the model needs. The G1 ships a RealSense depth camera, a Livox lidar and a microphone array, the kit of a robot expected to map a room. Figure 02 lists six RGB cameras, microphones and speakers, and no depth camera or lidar: the bet that a vision-language-action model can work out distance from pixels. Digit carries depth cameras and lidar for warehouse aisles. Of the 44 you can order, 10 list any sensor at all.

## The model: 78 name software, 19 name a model, GR00T is on none

The model turns sensor data into joint commands. 78 records say something here. 17 say reinforcement-learning locomotion, meaning the walking was trained in simulation. 13 point to open software. 7 say VLA. 19 name a specific behaviour model: Helix on Figure 02 and 03, GO-1 on [AgiBot A2](/robots/agibot-yuanzheng-a2/) and [Genie G1](/robots/agibot-genie-g1/), [Google DeepMind](/companies/google-deepmind/)'s Gemini Robotics on Apollo and the electric Atlas, GraspVLA on [Galbot G1](/robots/galbot-g1/), Redwood on NEO, Xpeng's VLA on Iron, ERA-42 on [Robotera L7](/robots/robotera-l7/), Carbon on Phoenix, Tesla's end-to-end networks on Optimus, Huawei Pangu on [Leju Kuavo](/robots/leju-kuavo/), Huisi Kaiwu on [Tiangong 2.0](/robots/tiangong-2/), Ant Group's model on [Robbyant R1](/robots/robbyant-r1/), an in-house VLA on [Menteebot](/robots/mentee-menteebot/) and Stanford's shadowing transformer on [HumanPlus](/robots/humanplus/). Nvidia's GR00T N1, released 18 March 2025 as an open humanoid foundation model and shown on NEO that day, is the named stack of no robot in our data.

Of those 19 machines with a named model, 5 say teleoperation yes, 8 partial, 3 no and 3 unknown. The stack line and the human line belong together.

| Machine | Named model | Teleoperation | Autonomy evidence in the record |
|---|---|---|---|
| [Figure 02](/robots/figure-02/) | Helix VLA, onboard | no; training data teleoperated | BMW Spartanburg sheet-metal loading, about 11 months in 2025, confirmed by BMW |
| [Figure 03](/robots/figure-03/) | Helix VLA | no; training uses teleoperation and human video | company videos of chores; independent reporting limited |
| [Galbot G1](/robots/galbot-g1/) | GraspVLA, GroceryVLA | no | autonomous picking in Beijing unmanned pharmacies, reported 2025 |
| [Atlas (electric)](/robots/boston-dynamics-atlas-electric/) | large behaviour models with TRI; Gemini Robotics | partial; demos autonomous, training data includes teleoperation | parts sequencing video October 2024; Hyundai plant testing reported |
| [Apollo](/robots/apptronik-apollo/) | Gemini Robotics plus Apptronik software | partial | Mercedes-Benz, GXO and Jabil pilots |
| [AgiBot A2](/robots/agibot-yuanzheng-a2/) | GO-1 VLA | partial; A2-D is a teleoperated data-collection variant | 106 km walk Suzhou to Shanghai, November 2025; factory logistics pilots |
| [Astribot S1](/robots/astribot-s1/) | VLA trained on teleoperated data | partial; launch video was a mix | household tasks in launch video; cooking demos 2025 |
| [Optimus Gen 2](/robots/tesla-optimus-gen-2/) | Tesla end-to-end networks | yes; We Robot units on 10 October 2024 remotely assisted | battery-cell sorting video; independent confirmation limited |
| [NEO](/robots/1x-neo/) | Redwood and 1X World Model | yes; Expert Mode lets 1X staff drive it | launch reviewers found most tasks teleoperated |
| [Phoenix](/robots/sanctuary-phoenix/) | Carbon, teleoperation-first | yes | Magna and Mark's pilots, company-reported |
| [Genie G1](/robots/agibot-genie-g1/) | GO-1 | yes; built for teleoperated data collection | table-top demos |
| [Xpeng Iron](/robots/xpeng-iron/) | Xpeng VLA and VLT | unknown | catwalk gait 2025; leg cut open on 6 November 2025 to show nobody was inside |
| [Optimus V3](/robots/tesla-optimus-v3/) | Tesla end-to-end networks | unknown | none |

A human can enter at three points. Live, through a controller or VR rig: the We Robot units and NEO's Expert Mode. In the data, where the model's training runs came from people in rigs; Figure, Boston Dynamics, AgiBot and Sanctuary all say so. And in the edit, where a video shows only the runs that worked. The records that say "no" and have someone other than the maker confirming the work are Figure 02, where BMW confirmed the pilot, and Galbot G1, where media reported the pharmacies. That is two.

## How to read any spec sheet

Start at the bottom. Actuation: is there a torque or payload number, or just the word electric? 173 of 178 have no torque figure. Joints: is the total given, and how many are in the hands? Compute: is the chip named, and can you log into it? Jetson with an SDK means yes; FSD-derived or Turing means no; "onboard computer" means ask. Sensors: is there anything beyond cameras, and does the hand feel? Model: is a model named, and does the same page say who collected its training data and how? Then the flag: does the maker state that no operator is involved in the demo, and has a customer, not the maker, confirmed a task done that way? Deciding between two machines, prefer the one whose sheet answers all six over the one with the better video. In our data a named computer is rarer than a machine you can order: 43 against 44.

## Sources

- [Unitree G1 product page](https://www.unitree.com/g1), Unitree
- [Helix: A Vision-Language-Action Model for Generalist Humanoid Control](https://www.figure.ai/news/helix), Figure AI
- [Farewell to HD Atlas](https://www.youtube.com/watch?v=-9EM5_VFlt8), Boston Dynamics (YouTube)
- [Xpeng newsroom](https://www.xpeng.com/news), Xpeng
- [NEO](https://www.1x.tech/neo), 1X
- [Tesla AI and Robotics](https://www.tesla.com/AI), Tesla
- [NVIDIA Isaac GR00T](https://developer.nvidia.com/isaac/gr00t), Nvidia
- [Galbot official site](https://www.galbot.com/), Galbot