One second inside a robot

You say "put the cup in the sink" and a humanoid does it in a few seconds. Inside those seconds, magnets push on coils 40,000 times a second, a learned policy writes joint targets 200 times a second, a language model re-reads the scene a few times a second, and the whole episode, including the part where it drops the cup, ends up in a training set weeks later. The robot works because each layer runs 5 to 50 times faster than the layer above it and trusts the layer below. It breaks where one layer is late, blind or out of its depth.

One request, seventeen steps

The walk-through below follows a humanoid built like a Figure 03 or an Optimus-class robot: a VLA brain split into a slow reasoning model and a fast policy, a learned balance layer, an EtherCAT line to every joint, and joint modules with a motor, a harmonic drive and two encoders each. Real robots differ in detail; the order of events and the time scales hold across the field. Press play or step through. The top half shows the robot; the bottom half lights the layer that is working, on a time axis where each tick is ten times the one before.

Twelve orders of magnitude

The single most useful picture of a robot is the time axis in FIG 14.1. Read it from the left. At 1 µs, EtherCAT's distributed clocks keep every joint in step. At 25 to 100 µs, each motor driver closes its current loop. At 1 ms, the whole-body layer balances the robot and the bus delivers a fresh set of targets. At 5 to 20 ms, the learned policy writes the next targets. At 100 ms to 1 s, the reasoning model looks again. Past a second, humans and clouds take over: the operator in seconds to minutes, the log upload in hours, the retrain in days or weeks.

The same axis tells you where money and failure sit. The left half, microseconds to milliseconds, lives in firmware on chips inside each joint. It is mature, cheap to run and made in China: a Unitree G1's entire compute subsystem costs about CNY 1,415 (US$200) estimate. The middle, 5 ms to 1 s, is where the frontier labs fight: Figure has committed US$3.5 billion of compute to Helix verified. The right half, seconds to weeks, is people, logistics and data, and it is the part a company in Southeast Asia can own without a chip team.

Figure's Helix 02 shows the boundary moving. Its S0 layer replaced 109,504 lines of hand-written C++ balance and whole-body code with a 10M-parameter network running at 1 kHz, trained across more than 200,000 parallel simulated environments verified. Learning has eaten down to the millisecond layer. The current loop and the bus below it stay classical, because physics at 40 kHz has no data problem to solve. See where learning replaces classical control.

The whole machine on one sheet

FIG 14.2 is the wiring diagram behind the walk-through. Every sensor ends at one of two computers. Every motor hangs off one EtherCAT line. One battery feeds everything, and one hardwired safety chain can cut motor power without asking any software.

Walk the diagram in the order a fault travels:

• Cameras to the main computer. Head and palm cameras stream images to a Jetson Thor-class computer drawing up to 130 W verified. NVIDIA raised the Thor T5000 module price from US$2,999 to US$4,999 at 1,000 units on 22 Jul 2026 verified, so compute is now one of the few humanoid parts getting dearer. Three brains and a cloud.
• Main computer to real-time controller. Joint targets go down over Ethernet at 50 to 200 Hz; robot state comes back up. The real-time controller runs PREEMPT_RT Linux or an RTOS, where a late answer counts as a wrong one. Real time.
• The EtherCAT line. One cable threads through every joint group and back. A cut or chafed cable in a moving arm takes out everything downstream of it, and Figure named the forearm, with its moving cables feeding the wrist motors, as Figure 02's top hardware failure point at BMW verified. Wiring and cable fatigue.
• Inside one joint. Driver, motor, reducer, two encoders, torque sensor, brake and bearing. In a rotary joint the reducer is about 36% of the cost, the torque sensor about 30% and the motor about 13.5% estimate. The bill of materials.
• Power and safety. A G1's 13-cell pack sits near 47 V nominal verified. The BMS reports charge and temperature on CAN. The emergency stop opens a contactor on the power bus. On a wheeled robot that is a safe state. On a legged one it means a fall, which is the unsolved problem in humanoid safety.
• Up to the cloud. Logs, maps and updates move over Wi-Fi or 5G to the maker's fleet cloud, which is also where a remote operator's takeover session comes in. That link is the main attack surface found in Unitree's robots. Cybersecurity.

From fleet to magnet

FIG 14.3 zooms in six times. Each panel is roughly ten to a hundred times faster and smaller than the one before, and each has its own specialists, its own suppliers and its own way to fail.

Each zoom level maps to a different kind of company. Magnets come from a supply chain China dominates and has licensed for export since April 2025, with a broader October 2025 expansion suspended until 10 Nov 2026 verified. Motors and drivers are a Chinese price war: Leadshine plans 2 million frameless motors in 2026 after delivering 120,000 in 2025 estimate. Joints are where Unitree makes its margin, since it builds its own joints, drivers and BMS verified. Arms and whole robots are the makers. Fleets are the customers, the distributors and the service firms.

Where the second goes wrong

Every step in FIG 14.1 has a characteristic failure, and each one shows up at a different distance from the robot. The list below runs from the magnet outward, which is also the order of how hard the fault is to fix in the field.

The failed attempt is the valuable part

Steps 14 to 17 are where the robotics industry's money is moving in 2026. The cup that tipped over, the operator who finished the job, and the synchronised record of both, are exactly the data every frontier lab now trains on. Physical Intelligence's π*0.6 learns from teleoperator interventions during autonomous runs, and π0.7 trains on "potentially suboptimal" autonomous data including failures, labelled with how well each attempt went verified. 1X guides NEO's launch autonomy at 60 to 70%, with a remote operator covering the rest verified.

Who captures that record decides who owns it. The robot writes it; the maker's fleet cloud receives it; the lab that trains the policy uses it. A third party sees it only if its contract gives it the logins and the logs, or if it operates the robots and the takeover desk itself. See why failures are the most valuable data and who holds the keys.

Terms in this chapter