Humanoid robots head to the factory floor, but the supply chain still runs through China

Global shipments jumped 432.1% year over year, the sharpest signal yet that humanoids have moved past the demo stage. IDC also raised its forecast for 2030 to more than 750,000 units, TechNode reported, citing IT Home.

Shipping robots and deploying them are different problems, though, and the newest reporting from the factory floor shows where the friction sits.

Data is the bottleneck

Innodata opened a motion-capture lab on 30 September in Ridgefield Park, N.J., to generate training data for humanoids and industrial robots. The company told The Robot Report that it captures 3D motion directly from bodies instead of inferring it from 2D video. It uses infrared optical tracking cameras from Vicon that measure movement to sub-millimeter accuracy.

"Physical AI has to earn its tokens one interaction at a time, and they have to be deliberate," Franklin Tanner, Innodata's vice president of robotics and physical AI, told The Robot Report. Tanner said the company's sensors are built to register the smallest motion of every joint, because "when you're training a humanoid that weighs almost 200 lb. [90.7 kg], your readings can't be in the ballpark."

Innodata's lab is one answer to a shortage that robotics teams call structural. No internet-scale corpus of physical interaction data exists, so every robot has to be taught deliberately.

Memory and sensing at the joint

On 1 October, SemiEngineering published a sponsor blog detailing the memory requirements inside humanoids. Distributed non-volatile memory handles boot management, telemetry logging and functional safety. Standard Flash and EEPROM, by contrast, introduce a 5 to 10 ms page-write delay. If a robot loses power inside that window, the active position log fractures and joint controllers lose absolute coordinates, which forces a mechanical homing sequence on reboot. The same piece cites Goldman Sachs analyst estimates of more than 6 million annual humanoid shipments by 2035, up from fewer than 15,000 in 2025.

Force and torque sensing is the other half of the problem. ForceN's founder and CEO, Robert Brooks, will give a talk on force and torque sensing for humanoids at RoboBusiness 2026 on 20 and 21 October in Santa Clara, The Robot Report said on 29 September. The company says it has supplied sensing systems to surgical, logistics and humanoid robots.

Tesla's hard road

The most detailed look at a single humanoid program came from The Information, as summarized by Ars Technica on 26 September. Tesla reportedly scaled Optimus production to hundreds of robots per week and is targeting more than 1,000 per week by the end of 2026. Workers still assemble hands and forearms by hand, and each one contains more than 100 small components. Touch sensors have proved unreliable enough that Tesla developed a glove-like sensor layer that can be replaced without swapping the whole hand.

Ars Technica also reported that some factory workers pushed back on wearing motion-capture suits, because they "knew the robots were designed to eventually replace them." Tesla moved data collection to dedicated teams and set up training hubs.

The supplier map is the part that does not change. Ars Technica noted Tesla's continued reliance on Chinese suppliers, and solar shows the same pattern: Electrek reported on 30 September that Tesla's planned Texas solar factory is caught in a US patent suit between Chinese equipment makers Linton Crystal Technologies and Zhejiang Jingsheng over crystal pullers, filed on 22 September in the Eastern District of Texas.

For anyone building a humanoid factory, the lesson from the past week is narrow. The robots are shipping in volume, mostly from China. The data, the sensors and the memory underneath them are still being figured out.