Humanoid robots are moving from demos to early commercialization—but today’s market is still defined less by “factory workers” and more by entertainment, education, and data collection. A teardown of Unitree’s G1 (a compact humanoid sold in multiple SKUs) provides a useful lens into what is structurally true about the sector: where costs really sit, what hardware is becoming commoditized, and why software and systems integration are likely to decide the winners.

IDC estimates global humanoid robot shipments reached nearly 18,000 units in 2025 (up 508% YoY), with a market size of about $440 million, concentrated in commercial performance, research/education, and data acquisition.

What is the Unitree G1, and why do teardowns matter?

Unitree’s G1 is a small, lightweight humanoid platform positioned for “high-dynamic” movement (running, jumping, performance-style motions) rather than continuous industrial labor. The G1’s importance is not that it is the most advanced robot in the world, but that it is a mass-producible reference design that reveals the industry’s current economics.

A teardown matters because it answers evergreen questions investors and operators keep asking:

• What does a humanoid robot really cost to build?
• Which components are defensible, and which are already “off the shelf”?
• What constraints stop humanoids from becoming factory machines?

How much does a G1 cost—and what does that imply for margins?

Pricing: one robot, many SKUs

The G1’s base version is priced around RMB 85,000 (after tax), while EDU versions range roughly RMB 169,000 to RMB 309,000, depending on compute, joint configuration (degrees of freedom), and dexterous hands.

This SKU strategy is structurally important: in early-stage robotics, vendors often use a common chassis and upsell high-margin options (compute modules, hands, sensors, warranties) to different customer types—schools, labs, integrators, and exhibitors.

Estimated BOM and gross margin

The teardown estimates the base unit’s full-robot BOM at ~RMB 41,600, with joint modules alone at about RMB 27,500. With estimated processing/manufacturing costs of about RMB 3,000, implied gross margin for the base version is around 40.7%.

For higher-end EDU versions, the report estimates gross margins rising to roughly 63%–67%, largely because incremental upgrades (more compute, higher torque, more DoF, dexterous hands) can be priced far above their marginal cost—especially when sold to institutions with budgets for R&D and demonstration.

Evergreen takeaway: in the near term, humanoid robot profitability is likely to come from platform + options, not from a single “perfect” all-purpose worker robot.

What drives the cost structure of a humanoid robot?

1) Joints are the biggest cost center

G1 has 23 degrees of freedom, and its economics are dominated by joint modules. The design integrates:

• motor + planetary gearbox + encoder + driver board (“four-in-one”)
• compact packaging with hollow-shaft internal wiring

This integration reduces assembly complexity, weight, and wiring failure points—but it also concentrates value into the joint supply chain.

2) Compute is increasingly modular

The base version uses a Rockchip RK3588, while the EDU configuration adds an Nvidia Jetson Orin NX (~100 TOPS) module for edge AI.

This reflects a broader trend: robot compute stacks are converging toward “developer kit” modularity, similar to drones and autonomous vehicles. As that happens, differentiation shifts away from the mere presence of compute and toward how well teams use it for motion control, perception, and reliability.

3) Sensors are often mature, purchasable parts

The head sensor package uses mainstream components such as:

• Intel RealSense D435i depth camera
• DJI Livox MID360 3D LiDAR

As more robots adopt similar perception hardware, the bottleneck becomes not “having sensors,” but sensor fusion, latency control, and robust autonomy in messy real-world environments.

How does the G1’s architecture work (in plain English)?

A typical humanoid control loop looks like this:

1. Perception layer: LiDAR and depth camera capture 3D geometry and close-range details. High-bandwidth data flows directly to the main compute (not over low-speed buses).
2. Decision layer: the main compute runs motion planning, balance control, and obstacle logic (the teardown describes reinforcement-learning-style control and a “world model” approach).
3. Execution layer: targets (joint angles/torques) are sent over a CAN bus daisy chain to each joint driver, which closes the loop using dual encoders for precise feedback.

Why this matters: humanoids are real-time systems. Small advantages in control algorithms, tuning, and timing often matter more than small differences in sensors or chips.

What are the key constraints preventing “industrial humanoids” today?

Battery and runtime are hard limits

G1 uses a 13-series lithium pack of about 0.42 kWh and an estimated 1–2 hours of runtime. That is broadly consistent with the power reality of legged locomotion.

For industrial settings, “1–2 hours” is usually not acceptable. It forces operational workarounds: battery swapping, frequent cool-down, or limited duty cycles.

Thermal management reveals product positioning

The teardown characterizes G1’s thermal approach as conservative: mostly passive cooling, with active cooling concentrated around the main controller and certain joints. The system also uses thermal derating when temperatures rise.

Structural point: thermal design is not just engineering—it is a product decision. Robots designed for demos, rentals, and education can tolerate short duty cycles; robots for factories cannot.

Payload is the gating factor for real work

Despite a 35 kg body weight, the G1’s single-arm payload is about 2 kg (higher in upgraded versions). Low payload limits tool use, handling, and the ability to mount heavier end-effectors (including many high-DoF dexterous hands).

Industrial implication: until humanoids can reliably lift and manipulate meaningful loads for long shifts, many “factory humanoid” narratives will remain pilot projects.

Is humanoid robot hardware becoming commoditized?

The teardown suggests many core items—sensors, compute components, memory/storage, and several mechanical parts—are mature and sourceable. Even where companies claim self-developed parts (motors, drivers), the broader ecosystem is rapidly learning how to produce comparable modules.

This is why many teams emphasize a view that can be summarized as:

• Hardware is increasingly reproducible
• Software and control are harder to copy

That does not mean hardware doesn’t matter; it means durable advantage may come from system-level integration: motion control, safety, reliability, calibration workflows, manufacturing quality, and developer tooling.

Who are the “real” players: robot brands or supply chains?

In humanoids, the “player map” has two layers:

1) Platform companies (robot brands)

They integrate the full stack: mechanical design, electronics, control algorithms, supply chain, assembly, and go-to-market. Over time, platform companies tend to consolidate because scale improves costs, reliability, and software ecosystems.

2) Component specialists (the picks-and-shovels)

A humanoid BOM concentrates spend in joints, reducers, motors, bearings, sensors, and lightweight structures. Even if robot brands consolidate, the upstream supply chain can remain diverse—though it may also consolidate around “preferred” vendors that meet reliability and volume requirements.

Why the market may end up with only a few major platforms: humanoids require high upfront R&D, safety validation, and continual software iteration. Once a platform reaches scale, it can spread costs across more units, attract developers, and negotiate component pricing—advantages smaller rivals struggle to match.

What might happen next in humanoid robots?

Expect faster progress in “short-duty-cycle” use cases

Over the next few years, growth is likely to remain strongest in:

• education and research platforms
• commercial exhibitions and performance
• teleoperation and data collection for training models

These applications fit today’s constraints (battery, thermals, payload) and monetize features like programmability and stable locomotion.

Industrial humanoids will require different joints

The teardown argues that purely rotary joint architectures face limits in torque density, stiffness, and sustained output under the weight constraints of a humanoid body. One proposed direction is linear-actuated joints (e.g., screw-driven mechanisms) for higher thrust density and self-locking behavior.

Whether this becomes the dominant path is still uncertain, but the underlying point is evergreen: industrialization will require architecture changes, not just incremental tuning.

Software will capture more of the value

As sensors and compute become standardized, differentiation will likely shift further toward:

• high-dynamic motion control and stability
• autonomy stacks that work outside lab conditions
• developer platforms, simulation, and data pipelines
• reliability engineering (MTBF, maintainability, calibration)

Related Coverage:

Unitree Robotics Founder Predicts Embodied AI’s 'ChatGPT Moment' Is Still Years Away

Unitree’s STAR Market IPO Filing Puts China’s Humanoid Robot Economics Under a Spotlight