RobotGuide

Look at real robots

Six companies, six architecture routes

Topic lists do not show you how the pieces assemble. These six teardowns are the assembly drawings: what each company uses for brain, cerebellum, and body, and why. After reading, every topic has a place to live.

Memorize this division first (it beats any terminology)

The brain decides «where to go» — thinking for 1 second is fine. The cerebellum decides «how to move without falling» — it answers every millisecond. The body is muscle and bone; data is experience. The six companies differ in how they solve these four things.

Sourcing note

Everything below is compiled from each company's official blog, launch events, and open repositories (source type noted). The «our read» sections are explicit judgments, not facts. Robotics information iterates fast — defer to the latest official disclosure.

Tesla Optimus

Humanoid · automotive-grade supply chain route

It moved its carmaking assets straight over: chips built in-house, joints built in-house, suppliers all ready-made — the biggest difference between it and every startup.

Brain

In charge of 'seeing and thinking'. The visual approach reuses the car's setup (cameras only, no LiDAR), and the compute chip is adapted from a vehicle chip — no development from scratch, which is where it starts ahead.

Cerebellum

In charge of 'keeping the body steady'. Joint actuators are developed in-house; standing and walking rest on classical control methods. Official videos have shown yoga-like balancing moves.

Body

All actuators developed in-house; the fingers are pulled by thin cords — the same idea as human tendons, hence the dexterous hands. The battery pack comes straight from the automotive-grade supply chain: cheap and stable.

Data

Mainly collected via demo videos and human teleoperation. Officially disclosed training details are scarce; outsiders can only reverse-engineer how it was made from the videos.

Public disclosures
  • Official line: actuators and the hand design are fully in-house, targeting cost-controlled mass production (launch events and Musk's public X account)
  • Compute and vision reuse the FSD team's accumulated work — the biggest gap versus startups (public interview statements)
  • The 2024–2025 We, Robot events showed batches of prototypes, but no independent third-party test data exists (public fact)

Our read (judgment, not fact)

  • Judgment: its core advantage isn't algorithms but the supply chain — automotive-grade cost curves for motors, batteries, and wiring harnesses that startups can't catch in three years
  • Judgment: opacity is both risk and opportunity — outsiders can only reverse-engineer the architecture from videos, and that reverse-engineering ability is itself an interview topic
  • For learning: it proves the pure-vision + in-house actuator route works, and it means perception and motor-control talent stay scarce long-term
Matching topics

Figure Helix

Humanoid · on-device dual-system route

The company that explains 'brain and cerebellum' most clearly: a slow one handles 'seeing and thinking', a fast one handles 'keeping the body steady', and both run on the robot itself.

Brain

System 2: a vision-language model of roughly 7 billion parameters (official blog figures), responsible for understanding scenes and instructions and deciding 'what to do next' about 7–9 times per second.

Cerebellum

System 1: a small model of tens of millions of parameters outputting continuous upper-body actions 200 times per second — translating the brain's intent into commands the joints can execute.

Body

The Figure 02 full robot: multiple DoF + dexterous hands; two GPUs ride inside the robot's body — no cloud dependency, it keeps working offline.

Data

Trained on data collected by human teleoperation. Official demos show it tidying clutter in an ordinary home — picking up items it never saw in training, thanks to the model's common sense.

Public disclosures
  • February 2025 official blog disclosed the Helix dual-system architecture with parameter scale and frequency figures (the sole source of the numbers on this page)
  • Home-scenario demo: completing tasks with items absent from the training set, via the VLA's common-sense generalization (official video)
  • BMW production-line pilot collaboration (publicly announced by both parties)

Our read (judgment, not fact)

  • Judgment: Helix is a 'textbook-grade' architecture narrative — a slow, smart brain and a fast, simple cerebellum bridged by intent; understanding it matters more than memorizing parameters
  • Judgment: putting the large model on the robot's own GPU instead of relying on the cloud is what keeps its latency controllable and makes it sellable; cloud-route companies will all have to answer this question eventually
  • For learning: System 1's real-time control stack is exactly the D4/D7/B-domain content on this site — in the brain-cerebellum era, cerebellum engineers face no unemployment risk
Matching topics

Unitree

Quadruped → humanoid · cost-breakthrough route

The company that crushed prices: motors built in-house, the G1 officially priced from RMB 99,000; its walking gait was 'fallen' into existence with reinforcement learning in simulation.

Brain

Doesn't bet on large models. The main controller runs navigation and perception (LiDAR + camera, depending on the model), leaving the 'brain' slot to developers and partners.

Cerebellum

The reinforcement learning route: let the robot fall millions of times in simulation to forge a walking policy, then move it to the real robot — why it can backflip and survive a kick without falling.

Body

The foundation is in-house motors and reducers, and that's where the cost advantage comes from; the SDK and developer docs are open source, and Chinese university labs use it as a research platform.

Data

Mainly trained on simulation data (platforms like Isaac), without relying on large-scale human teleoperation.

Public disclosures
  • G1/H1 official prices and specs are public; the SDK and developer docs are open source, and a large share of Chinese universities' RL locomotion-control papers use its platform
  • The 2025 Spring Festival gala yangge dance and marathon events turned 'locomotion maturity' into a publicly visible metric
  • Official line: motors, reducers, and other core actuator parts are developed and produced in-house

Our read (judgment, not fact)

  • Judgment: Unitree proved that 'locomotion control + actuators' can stand alone as a business model, without waiting for large models to mature
  • Judgment: RL sim2real has gone from paper technique to the standard cerebellum solution at humanoid companies — for locomotion students, simulation-platform skill is the new baseline
  • For learning: it's the best hands-on platform for students (price + open SDK), and this site's F1/F2 topics map directly onto its technical route
Matching topics

Boston Dynamics Atlas

Humanoid · deepest control accumulation route

The ceiling of whole-body control: behind the parkour and backflips is a decade-plus of control technology accumulation. In 2024 it switched wholesale from hydraulic to electric.

Brain

Partnering with Toyota Research Institute since 2024 (announced by both sides), exploring plugging large models into the task-understanding layer — previously it could only execute motions, not 'understand tasks'.

Cerebellum

MPC + whole-body control (WBC) accumulation is the industry's deepest: balance, stepping, and arm pose managed simultaneously. Switching to electric actuators is like moving the control stack into a new house — it needs retuning.

Body

All-electric version since 2024 (official release): more power-efficient and easier to maintain than hydraulics, but the hydraulic-era explosiveness must be rebuilt.

Data

Mainly physics-model driven (not data driven); the public output is mostly official demo videos.

Public disclosures
  • April 2024 official announcement retiring the hydraulic version and launching the all-electric version (official blog, with farewell video)
  • The Toyota Research Institute collaboration is public on both companies' sites: the large model handles the task-understanding layer, Atlas's own control stack handles execution
  • Parkour, backflips, and similar abilities are all disclosed via official videos; third-party reproductions are rare

Our read (judgment, not fact)

  • Judgment: Atlas's value lies in the depth of its control-stack accumulation — something ten years can't buy; the electric version is like moving the old foundation into a new house
  • Judgment: even the company with the strongest control is plugging in large models, showing that 'missing brains' is an industry-wide consensus, not a gimmick
  • For learning: WBC / ZMP / impedance control and the other B-domain topics are the interview language of companies like Atlas
Matching topics

AgiBot

Humanoid · data and open-source route

The Chinese company doing the most on 'data + open source': it open-sourced a million real-robot trajectories — teams without a real-robot fleet can also train models on the data.

Brain

The GO-1 model (officially released 2025): a VLM handles 'seeing + hearing' while an action expert outputs continuous control signals.

Cerebellum

The Explorer series robots: a classical control stack at the base, with RL-based locomotion control on top (per public demos).

Body

Humanoid robots like the Explorer A2, plus the small Lingxi series (partly open-source hardware), priced low enough for university labs to afford.

Data

Biggest highlight: the AgiBot World dataset is open source (officially million-scale real-robot trajectories), among the largest Chinese real-robot manipulation datasets.

Public disclosures
  • Both GO-1 and the AgiBot World dataset were open-sourced on official channels (public information, 2025)
  • Lingxi-series pricing and open-source repos are public, letting Chinese universities reproduce at low cost
  • Mass-production progress and factory training scenarios have steady public coverage

Our read (judgment, not fact)

  • Judgment: open-sourcing the dataset is its biggest contribution to the industry — the VLA race is essentially a data race, and AgiBot World lets students without a real-robot fleet take part
  • Judgment: the ViLLA architecture represents the mainstream Chinese VLA approach: add an action expert on top of an open-source VLM rather than training a foundation model from scratch
  • For learning: the first step into VLA is downloading AgiBot World + OpenVLA and reproducing one training run — more useful than reading ten papers
Matching topics

UBTECH Walker S

Humanoid · factory-training route

The Chinese manufacturer that actually put humanoids on factory production lines for training. Conservative architecture: factories want identical results every time, not occasional brilliance.

Brain

Layered architecture: a large model orchestrates at the task layer while mature visual servoing handles execution — stable reproduction first, no end-to-end bet.

Cerebellum

Servo drives developed in-house (per company public materials), with the focus on steady motion and precise operation.

Body

Walker S is designed to industrial standards: dual-arm operation + mobile base, with payload and precision benchmarked to production-line requirements.

Data

Factory training generates real task data (per public reports from partners) — the biggest difference from pure-lab vendors.

Public disclosures
  • Training collaborations with multiple automakers / contract manufacturers are publicly reported (on-site sorting, inspection, and similar scenarios)
  • Servo drives and joint solutions developed in-house (per company public technical materials)
  • As a listed company its public disclosures are relatively complete — the best sample for watching 'how humanoids enter factories' in China

Our read (judgment, not fact)

  • Judgment: factories are the only scenario where humanoid robots can close the business loop today — its route is conservative but the cash flow is healthy
  • Judgment: layered architecture is more credible than end-to-end in industrial settings: the production line wants stable reproduction, not clever generalization
  • For learning: roles like system integration and on-site commissioning (F6) are exactly where companies like Walker hire the most
Matching topics

After the teardowns, go back to the topic map for depth, or to brain and cerebellum for where all six routes converge next.