Humanity's Last Machine: A Deep Dive on Humanoid Hardware

TL;DR

An interactive long-form report on the humanoid robot hardware stack — components, landscape of startups, key suppliers, and US/China geopolitics. Authored by Sourish Jasti, Zoey Tang, Intel Chen, and Vishnu Mano (with RoboStrategy support) after ~8 weeks of factory visits, expert interviews, and BOM analysis. Frames the humanoid bill of materials as a complex-system supply-chain study and asks which parts are structurally expensive vs compressible with scale, and where the US and Chinese ecosystems diverge on supply chain, industrial density, and capital flow. Relevant to Luma researchers as physical context for the VLA / world-model / robot-policy work filed on this wiki — the “what is being trained to move” side of the stack.

Key claims

The report is organized around four axes: hardware components, startup landscape, key suppliers, and geopolitics [Site navigation].
Its stated purpose is not to predict a winning humanoid design but to give first-principles intuition for reading any humanoid — prototype or production — and inferring what may break, be hard to manufacture, be cheap at scale, or remain expensive [Report introduction, cited via Adafruit summary].
Actuators and joint assemblies are named as the largest and least scalable cost line in the BOM, with a typical humanoid using roughly 25–30 actuators [Hardware Components section, cited via Hardware FYI summary].
The US and Chinese ecosystems are compared along three axes — supply chain, industrial density, and capital flow — to explain structural advantages, bottlenecks, and where durable value capture is plausible [Geopolitics section, cited via Adafruit summary].
The site is presented as inseparable from the report (an interactive scroll-driven artifact) with a PDF version linked for offline reading [Site itself].

Method

The authors ran an ~8-week research process combining factory and company visits, interviews with industry experts, and collaboration with contributors across the humanoid stack. The output is a scroll-driven website that walks from individual mechanical components outward through joint assemblies, sensing, compute, then out to the map of startups, the tier-1/2/3 supplier graph, and finally the US-vs-China industrial-density and capital-flow comparison. A PDF mirror exists for offline reading. The design goal is a reference document rather than a prediction — a shared vocabulary and BOM-level intuition that any reader (investor, engineer, roboticist) can apply to inspect a specific humanoid design.

Results

Not a benchmark artifact — the “results” are qualitative structural findings:

Actuation dominates cost and scaling. With ~25–30 actuators per humanoid, actuator and joint-assembly cost is the largest BOM line and the least compressible with volume, making it the primary bottleneck for cost curves.
Compressible vs structural cost split. The framework distinguishes BOM lines that fall on volume (electronics, structural machining) from those that stay expensive (precision actuators, harmonic drives, high-torque motors, sensing).
US/China divergence. Chinese ecosystems win on supply-chain density and vertical integration at the actuator tier; US ecosystems win on capital availability and end-integration. Neither has a clear geopolitical winner yet — general-purpose robotics is called out as one of the rare frontiers where the two started at roughly the same time.

Why it’s interesting

Almost every VLA and world-model paper filed on this wiki — Hy-Embodied-0.5-VLA: From Vision-Language-Action Models to a Real-World Robot Learning Stack, Geometric Action Model for Robot Policy Learning, π0.7: A Steerable Robotic Foundation Model with Emergent Compositional Generalization, Flexion Reflect v1.0: The Path Towards Long-Horizon Autonomous Humanoid Work — implicitly assumes a humanoid or bimanual manipulator as its embodiment, but the wiki has almost no filed material on what that hardware actually costs, where its parts come from, or which parts are the load-bearing bottleneck. This report is the counterpart to those pages: the physical side of the software stack that dominates most of the robotics literature the team files. It also contextualizes the Unitree open-source data drop (Unitree open-sources UnifoLM-WBT-Dataset — humanoid whole-body teleoperation dataset) and functional-safety scaffolding like Inside NVIDIA Halos for Robotics: A Full-Stack Functional Safety System for Physical AI — both of which assume a certain hardware baseline (~25-30 DOF humanoid) that this report makes explicit. Contrasts with the “hobby-tier” side of the wiki like Fauna Sprout: A lightweight, approachable, developer-ready humanoid robot, which explicitly optimizes for approachability by dropping actuator count.