Compute, wiring and power

A humanoid is a battery of about 1 to 2.3 kWh driving dozens of motors, with a stack of computers deciding what each motor does thousands of times a second. The battery sets how long it works, heat sets how hard it can work, and the wiring decides how often it breaks. The main AI computer got up to 101% more expensive in July 2026 while almost every other part got cheaper, and the battery rules decide whether you can put a robot on a plane.

Three brains and a cloud

No robot runs on one computer. It runs on a hierarchy, fast and simple at the bottom, slow and clever at the top. Chinese engineers describe the top two layers as the 大脑 dànǎo ("big brain", the planner and AI model) and the 小脑 xiǎonǎo ("cerebellum", the motion controller). Both words are standard in Chinese robotics.

• Joint microcontrollers sit on each motor driver board and run the motor's current loop at 10 to 40 kHz verified. They are cheap chips with no operating system or a tiny real-time one.
• The real-time controller runs balance and whole-body control at 500 Hz to 1 kHz. Figure's Helix 02 runs a 10-million-parameter network here at 1 kHz, and it replaced 109,504 lines of hand-written C++ verified Jan 2026.
• The main controller, the AI computer, runs perception and the robot's AI model: a slow reasoning model at 1 to 10 Hz and a fast visuomotor policy at 50 to 200 Hz. This is where NVIDIA's Jetson modules live.
• The cloud trains models, stores maps and logs, and pushes updates. It is too slow and too unreliable for control; a robot must stay safe when the network drops.

The chips, and what they cost

Chip makers quote AI performance in TOPS, trillions of operations per second. Treat every TOPS number with suspicion: NVIDIA quotes Jetson Thor at 2,070 TFLOPS in FP4, a 4-bit number format, while Horizon quotes its chip in INT8 TOPS, and the two are not comparable without knowing the format. What matters is how fast your model actually runs at the power you can afford.

NVIDIA raised Jetson prices by 50 to 101% on 22 Jul 2026 with no public explanation; observers blame memory prices verified, CNX Software. The result is odd: in a Unitree G1 base, the whole compute subsystem costs about CNY 1,415 (US$200) estimate, while a Thor module for a Western humanoid now costs US$4,999. The compute line is one of the few in a humanoid bill of materials going up, and it is a reason Chinese makers are testing domestic chips from Horizon Robotics and Rockchip.

Real time: late is wrong

A real-time system is one where an answer that arrives late is a wrong answer. A balance controller that computes the perfect torque 5 ms late may drop the robot.

Buses: the robot's spinal cord

A bus is a shared wire that many devices talk over, like a party line. Running a separate cable from the controller to each of 30 to 50 joints would weigh too much and break too often, so joints hang off one or two buses.

• EtherCAT (Beckhoff, 2003) is the standard bus for multi-axis robots: about 100 Mb/s, with distributed clocks that keep every joint synchronised to under a microsecond verified. Its trick is processing on the fly, shown below.
• CAN bus (Bosch, 1986) runs at 1 Mb/s, and devices take turns by arbitration: the message with the higher priority wins the wire. CAN FD raises that to 5 to 8 Mb/s verified. It is cheap, robust and everywhere in cars, and small humanoids and quadrupeds often use it for their joints.
• RS-485 is an older, cheaper serial link over a twisted pair, common in hobby and low-cost servos.
• Ethernet connects the main computers to each other and to high-bandwidth sensors. Orbbec's Gemini 435Le and RealSense's D555 depth cameras send images over Power over Ethernet verified; other cameras use USB or dedicated serial camera links.

Wiring, slip rings and cable fatigue

The wiring harness is every cable in the robot bundled, routed and connected: battery power to each joint, bus lines, camera cables, emergency stop circuits. In a humanoid the harness has to cross every joint, and every joint crossing flexes the cable millions of times. Copper that flexes fatigues and cracks, connectors loosen, and the fault is intermittent, the hardest kind to find.

The evidence is concrete. Figure says the forearm was Figure 02's top hardware failure point during its 11 months at BMW: tight packaging, a three-axis wrist, heat, and a microcontroller distribution board with moving cables feeding the wrist motors. Figure 03 removed the board and the moving cables, so each wrist motor controller now talks straight to the main computer verified Nov 2025. Unitree routes wires through hollow joint shafts verified, which protects them from snagging and limits how far the joint can twist.

A joint that must turn without limit needs a slip ring: brushes riding on rotating metal rings that carry power and data across a spinning joint. Boston Dynamics' production Atlas has joints with continuous 360-degree rotation verified; how it passes power through them is not public. Slip rings wear and add electrical noise, so they are a maintenance item.

Batteries

A battery pack is many battery cells wired in series to raise the voltage and in parallel to raise the capacity, plus a management computer, wiring, fuses and a case. The cells are where the chemistry lives; the pack is where the engineering lives.

Two chemistries

Figures from the energy primer, ±20% estimate. Humanoids mostly use NMC in cylindrical 18650 or 21700 cells, because every kilogram of battery is a kilogram the legs must carry estimate. LFP is safer and lasts longer, and it suits wheeled robots that can afford the weight. Nominal cell voltage is about 3.6 to 3.7 V for NMC and 3.2 V for LFP, so a 13-cell series string like the G1's sits near 47 V nominal.

What named robots carry

Humanoid runtime clusters at 2 to 5 hours, against an 8 to 12 hour shift for a wheeled delivery robot. The arithmetic is the reason: a walking humanoid draws hundreds of watts just to stay upright, and a tray robot rolling on a flat floor draws a fraction of that.

Peak power and the battery management system

Average draw sets runtime; peak draw sets whether the robot survives a hard move. A humanoid standing up, jumping or catching a fall pulls a spike of current many times its average. Unitree's own maintainers confirm that a large instantaneous joint torque when a G1 switches to a new control policy can trip the battery's over-current protection and shut the robot down verified. The cells, the wiring and the fuses must all carry that spike.

Batteries age faster when drained deep. Pudu's manuals say running below 20% impairs battery life and below 2% the robot locks out until charged, and they tell venues to charge stored robots every 3 months and store them above 60% verified. A tired battery is one of the most common service calls on a wheeled robot fleet.

Swap, dock or charge in place

Robots escape the runtime limit in four ways, and each one shapes how a site runs them.

• Battery swap. UBTech's Walker S2 walks to a rack and swaps its own pack in about 3 minutes, and decides for itself whether to swap or charge verified. Atlas swaps its own battery too; Apollo and the Unitree H1 take a hand-swapped pack.
• Fast charging. Digit 5 recharges in 9 minutes at high current, for a 10:1 run-to-charge ratio and 20+ productive hours a day preliminary. NEO quick-charges 6 minutes per hour of runtime verified.
• Wireless charging. Figure 03 has coils in its feet and charges at 2 kW when it steps onto a pad, with cooling built into the die-cast battery housing verified.
• Charging dock. The standard for tray and cleaning robots: the robot drives onto contacts between jobs. Pudu's ET1 scrubber has an 8-in-1 dock that also refills and drains water verified.

Where the watts go

Power consumption in a walking humanoid is dominated by the joints. Nobody publishes a full power budget, so the diagram below is an illustration built from the few numbers that are public: about 500 W walking for an Optimus-class robot unverified, up to 130 W for a Thor-class computer verified, and 10 W for a Unitree L2 lidar verified.

Heat and derating

Heat is the hidden ceiling on what a robot can do. A motor's torque is proportional to its current, and its copper loss is proportional to current squared. Peak torque is set by magnetics; continuous torque is set by how fast the joint sheds heat, roughly scaling with the square root of the heat it can dissipate energy primer. Spectacular joint torque figures on a spec sheet are usually 5-second numbers.

When a joint gets too hot, its driver cuts the current it allows, which is called derating, and the robot gets weaker or stops. The most concrete public case: running NVIDIA's GR00T whole-body controller on a Unitree G1, both ankle-roll motors hit the 90 °C warning within 5 to 10 minutes, even standing idle, while Unitree's stock controller keeps the same ankles at 30 to 45 °C verified, owner report. Same hardware, a 45 to 60 °C difference, caused by software.

Thermal management climbs a ladder. Passive fins and the metal body come first. Unitree uses local air cooling with fans verified. Liquid cooling comes next: water carries about 4 times the heat per degree of air and conducts it about 24 times better energy primer, and Figure built cooling into its battery housing for 2 kW charging verified. The compute module is the other hot spot: 130 W in a sealed head or torso needs a real heat path.

Water, dust and interference

An IP rating is a two-digit code for how well a case keeps out solids and water. The first digit covers dust (6 is dust-tight), the second covers water (4 is splashing, 7 is brief immersion). Atlas and the Unitree B2 are rated IP67 verified. The Unitree A2 is IP56 per a reseller unverified. A Pudu KettyBot is IP20, which means no water protection at all, and Pudu says its robots "cannot be soaked in water, but light splashing is allowed" verified. Only NEO's hands are waterproof; the body is moisture-sensitive unverified. Water and oil ingress are standard warranty exclusions, so the IP rating tells you which sites a robot can work in without voiding its cover.

EMC (electromagnetic compatibility) means the robot neither jams nearby radios nor malfunctions near strong fields. Dozens of motor drivers switching tens of amps at 10 to 40 kHz make a robot a strong radio emitter. Testing labs such as SGS test EMC alongside electrical, mechanical and functional safety verified, Pudu's BG1 scrubber carries an FCC Class A rating verified, and Unitree's warranty excludes damage from strong electromagnetic interference verified.

Shipping lithium: UN38.3 and dangerous goods

Every lithium cell and pack must pass UN38.3, the UN transport test series, before it can be shipped. Figure's Figure 03 pack is UN38.3 certified and targeting UL certification verified. Lithium batteries travel as dangerous goods, with packing, labelling and paperwork rules that differ by air, sea and road.

The practical consequence is size. The airline carry-on limit for a lithium battery is 100 Wh energy primer. A G1 pack is about 420 Wh and a Figure 03 pack is 2,300 Wh, 4 to 23 times that limit, so robot batteries go as declared cargo, never in your luggage, and a sample robot from Shenzhen cannot come home with you on the plane. Three things follow for importing robots and spares into Southeast Asia:

• Every battery pack model needs its own UN38.3 test summary. Without it, freight forwarders refuse the shipment.
• Spare batteries are the slowest and most expensive spare part to move. Hold stock locally and plan lead times around sea freight.
• Check the current carrier and country rules with your freight forwarder before each first shipment; they change, and they are stricter for loose batteries than for batteries installed in a robot.

Connectivity, firmware and OTA

Robots talk to the outside world over Wi-Fi, 5G and Bluetooth; NEO has all three verified. Figure 03 adds a 10 Gbps mmWave radio to offload training data from the fleet verified. Connectivity is a real deployment cost: BMW improved 5G coverage in its hall for Figure's robots verified, and Pudu lists "loss of signal" as a standard fault verified.

Firmware is the software baked into each chip: the motor drivers, the BMS, the sensors. An OTA update sends new software and AI model weights over the network. Good designs keep two copies of the system, A and B, so a bad update rolls back to the last working one design pattern.

Terms in this chapter