Humanoid robot factories scale up, but the data still has to be earned

The newest piece of the humanoid robot factory stack is not a robot. It is a lab. Innodata said on 30 September that it has opened a facility to capture 3D motion data directly from human and mechanical bodies, in an effort to feed the models that drive humanoids and industrial robots. The Ridgefield Park, N.J. company, founded in 1988, framed the move as a fix for a shortage it says every robotics team hits.

According to The Robot Report, Innodata claims most data providers infer 3D motion from 2D video, while its sensors register the movement of every joint. Franklin Tanner, the company's vice president of robotics and physical AI, told the outlet that a 2D grid of pixels inevitably introduces mistakes. He also gave a number for the stakes: a humanoid can weigh almost 200 lb (90.7 kg), so readings cannot be "in the ballpark." The lab will also validate performance figures that robots generate internally. That is a quieter but sharper point. As humanoid factories multiply, the numbers attached to them, cycle times, success rates, throughput, are increasingly self-reported by the companies selling the machines.

Factories, orders and a Texas lawsuit

The same day, two funding and legal stories showed how the supply chain around those factories is being assembled. Tesla's planned solar factory in Texas has been dragged into a US patent lawsuit between two Chinese equipment makers, Electrek reported on 30 September. Linton Crystal Technologies, a Rochester, New York company wholly owned by China's Dalian Linton, sued Zhejiang Jingsheng in the Eastern District of Texas on 22 September, asserting two patents covering part of a Czochralski puller. Jingsheng, valued at about 53 billion yuan, disclosed the suit to investors on 24 September and says its products use "technical solutions entirely different" from the patents. It also says it has not been formally served and expects no material impact. Chinese industry reports say Jingsheng won a roughly 3 billion yuan (about $420 million) Tesla order for crystal pullers earlier this year, though neither Tesla nor Jingsheng has confirmed that. Tesla is not a party to the case. Linton is asking for damages, treble damages and a permanent injunction.

The injunction request is the part that matters for Tesla, which filed in August for a $10.1 billion "Project Crystal Sun" factory in Fort Bend County with production planned for early 2029. Ingot pulling is the first step in that chain. No pullers, no wafers, no cells.

In Europe, Paris-based Inbolt raised 11 million euros to push real-time vision and control software into industrial robots. Tech.eu reported on 30 September that the round was led by Shift4Good, with Bridges Climate Transition Partners and existing investors BNP Paribas Développement and Ora Global. Total funding now stands at 30 million euros. Inbolt says its system runs on more than 200 robots in over 100 factories across three continents, with customers including Bosch, Beko, Flex, Ford, Stellantis and Toyota.

The speed problem

Speed is the other wall. MindOn published a technical post on 30 September introducing Mind-1, a physical AI model it says cuts inference latency in robot manipulation from 82ms to 32ms. The company argues that many embodied systems can complete complex tasks but at a fraction of human speed, which makes them hard to slot into real workflows. Its post names the constraint precisely: at an end-effector speed of 3 metres per second, 10ms of latency equals roughly 3cm of movement. That is enough to spoil a precise grasp or insertion. MindOn also blames training data, much of which comes from teleoperation that is slower than natural human motion, teaching models a conservative tempo.

Independent evaluation suggests the gap is not closing evenly. Fig Inc. published item-level results on 30 September across web and physical benchmarks, including VLABench, Bench2Drive, IndEgo and Assembly101. Its finding is blunt: in every model pair tested, the lower-scoring model solved at least one task the higher-scoring one failed. A benchmark average, the authors write, hides how much a model's success varies across tasks. That matters for anyone buying a humanoid for a factory line. A single headline score does not tell a plant manager which tasks the machine will fumble.

What is actually shipping

Destro AI came out of stealth on 30 September with an $8 million seed round, and its pitch is a useful counterweight to the humanoid factory narrative. TechCrunch reported that the startup does not build robots. It builds an intelligence layer that directs robots and human workers together in logistics settings, using three cart-moving robots from Miva Robotics at a Yusen Logistics facility in the Pacific Northwest. Yusen's Richard Brunelle told TechCrunch he asked Destro to adapt its picking and packing tool for cross-docking, a task he said nobody was addressing. Destro founder Manthan Pawar told the outlet the company is on a path to cash-flow positive by the end of this year.

Then there is the harder evidence from the maker side. Rongzhong Li, writing on Hackster on 30 September, says Petoi has shipped more than 30,000 quadrupeds since its 2018 Nybble campaign, and that its third product, Quaddle, is on Kickstarter from $99. His account of ten years of shipping robots is a reminder that volume in robotics has so far come from small, cheap, limited machines, not from bipeds.

KIDZ AI, a Nasdaq-listed company worth about $2.4 million, said on 30 September that it sent Tesla a procurement inquiry for 500 Robotaxis. Electrek noted that the inquiry is not a binding order, that the plan is in a preliminary phase, and that 500 Cybercabs at Elon Musk's own "under $30,000" target would cost about $15 million, roughly six times the company's market value. Its stock fell about 6% after opening.

Read together, the 30 September cluster points to a factory buildout that is real but lopsided. The robots are coming. The data, the speed and the honest numbers are still being negotiated.