Physical AI: Hardware, Economics, and the Humanoid Race — A Detailed Breakdown of Part Two

The humanoid form of robots is advantageous because the entire world is designed for humans: there is no need to redesign the environment, and training data is easier to obtain through teleoperation and motion capture. However, half of the actually working 'general-purpose' robots are torsos on wheeled bases (for example, China's Galbot). The main problems with humanoids are: actuators (harmonic drives, quasi-direct drive QDD, planetary roller screws—each school has its own trade-offs), hands (20+ degrees of freedom and tactile sensing require complex engineering; Tesla, 1X, and Figure are actively improving them), batteries (typical runtime is 2–4 hours; UBTech taught the Walker S2 to autonomously swap its battery). Price range: China's Unitree G1—about $16,000, the lower-end R1—less than $6,000; Western industrial models—hundreds of thousands of dollars. Current practical deployments are focused on warehouses and factory logistics: Agility Robotics (Digit) works at GXO's warehouse under a RaaS model, Figure assists BMW (30,000 car bodies assembled by 2025; from 2026, Figure 03 with a sequencing task), Apptronik (Apollo) at Mercedes-Benz's plant, Boston Dynamics (Atlas) is being tested at Hyundai's factories. So far, the count of robots is in the single digits, at most hundreds, but contracts are being renewed. Home robots (1X NEO for $20,000 or $499 per month) perform complex tasks in Expert Mode under remote operator control, which shows the unreadiness of autonomy for the home: slow speed, high cost of error, and privacy concerns. Chinese robots are cheaper than Western ones, and this determines whose hardware the world learns on.