Robots Get Better Hands, and a Legal Fight Over Who Builds Their Solar Furnaces

Innodata, a data engineering company founded in 1988 and based in Ridgefield Park, New Jersey, said the new research and development facility will capture 3D motion directly from human and mechanical bodies, rather than inferring it from 2D video. The Robot Report reported the launch on 30 September.

The pitch is simple: robots do not have an internet to learn from. Large language models scraped a web that already existed. Humanoids have to generate their own corpus, one movement at a time, and that is slow and expensive.

"Physical AI is growing faster than any other segment in AI, but every robotics team hits the same wall: There isn't enough real-world interaction data, and what exists is expensive and slow to produce," Innodata CEO Rahul Singhal said in a statement carried by The Robot Report. Franklin Tanner, the company's vice president of robotics and physical AI, told the same outlet that its sensors are built to register the smallest motion of every joint, and that a humanoid weighing almost 200 lb (90.7 kg) leaves no room for ballpark readings.

The lab will also independently validate performance data that robots generate internally, according to Innodata. That second function matters as much as the first. A robot's own logs are not a neutral record, and companies deploying them in factories need something closer to an audit than a self-report.

Latency, not just capability

While Innodata attacks the data problem, a startup called MindOn published a technical post on 30 September claiming its Mind-1 model cuts robot inference latency from 82 milliseconds to 32 milliseconds and reaches human-level cycle times on a set of logistics and everyday manipulation tasks. The company says some tasks are completed faster than a human would.

The post is unusually specific about why speed is hard. At an end-effector speed of 3 metres per second, 10 milliseconds of latency equals roughly 3 centimetres of travel, enough to spoil precise grasping or insertion. MindOn also blames training data: much robot manipulation data comes from teleoperation, which is slower than natural human motion, so models learn a sluggish tempo along with the task.

"A robot that can perform a task but takes five or ten times longer than a human is still difficult to integrate into real workflows," the company writes.

Those numbers are MindOn's own and have not been independently reproduced. Treat them as a vendor claim with a plausible mechanism, not a settled result.

Independent evaluation work published the same day suggests why single headline numbers deserve suspicion. Fig Inc released a technical report and a dataset called RIDGE, covering seven models on VisualWebArena and three models across four physical benchmarks: Bench2Drive, VLABench, IndEgo and Assembly101. Its central finding is that benchmark averages hide how unevenly models perform.

In every model pair the authors tested, the lower-scoring model solved at least one task the higher-scoring one failed. A model update could leave the average almost unchanged while flipping many individual tasks, and an update that raised the average still lost tasks the older version handled. The report covers Gemini 3.5-Flash, Astra, Opus 5.5 and others. For anyone choosing a robot stack, the practical lesson is that a leaderboard rank tells you less than the specific task category you plan to run.

Money moving into perception

Paris-based Inbolt raised 11 million euros to push real-time vision and adaptive control into industrial robots, Tech.eu reported on 30 September. The round was led by Shift4Good, with Bridges Climate Transition Partners and existing investors BNP Paribas Développement and Ora Global participating, bringing total funding to 30 million euros.

Inbolt targets a specific failure mode: even in highly automated factories, robots stall or produce defective parts when a component arrives misaligned or tooling wears. The company says its 3D vision plus hardware-agnostic AI layer is deployed on more than 200 robots in over 100 factories across three continents, with customers including Bosch, Beko, Flex, Ford, Stellantis and Toyota. Those deployment figures come from the company and its investor announcement.

On the same day, Destro came out of stealth with an 8 million dollar seed round, TechCrunch reported. The startup does not build robots. It builds a coordination layer that tells both machines and human workers what to do, and it is running a pilot at a Yusen Logistics facility in the Pacific Northwest with three cart-moving robots from Miva Robotics.

Manthan Pawar, Destro's founder, told TechCrunch that not being a robotics company is a competitive advantage. Richard Brunelle, director of automation for Yusen Logistics' American arm, said he asked Destro to adapt its picking and packing tool to cross-docking, where goods move from one truck into mixed loads for other trucks.

Two funding rounds and one product launch in a single day is not a boom by itself. But the pattern across all three is consistent: the money is going into data, perception and coordination, not into yet another bipedal chassis.

A patent suit lands on Tesla's solar plan

The most consequential item of the week is not a robot at all. Linton Crystal Technologies, a Rochester, New York company wholly owned by China's Dalian Linton, sued Zhejiang Jingsheng in the US District Court for the Eastern District of Texas on 22 September, Electrek reported on 30 September. The patents, No. 11,255,024 and No. 11,814,746, cover a seed lifting and rotating system for Czochralski crystal pullers, the furnaces that draw monocrystalline silicon ingots from molten silicon.

Jingsheng disclosed the suit to investors on 24 September, saying its products use technical solutions entirely different from the patents, that it has not been formally served, and that it expects no material impact. Linton is seeking damages, treble damages for willful infringement and a permanent injunction.

The injunction is where Tesla enters, indirectly. Chinese industry reports say Jingsheng won a Tesla order worth roughly 3 billion yuan, about 420 million dollars, for crystal pullers earlier this year, according to Electrek. Tesla is not a party to the case, and neither Tesla nor Jingsheng has confirmed the order.

The timeline reported by Chinese trade press runs from a Tesla tender for 210mm monocrystalline pullers, wafer cutting equipment and quartz crucibles that closed in February, through contracts signed in March, shipments starting in April and full delivery planned for the third quarter. In August, Tesla filed for a 10.1 billion dollar factory in Fort Bend County under the name Project Crystal Sun, covering ingot growth through finished modules, with production planned for early 2029. The Robot Report's sister outlet Electrek has been tracking the buildout since March, when Tesla was reported to be in talks to buy 2.9 billion dollars of Chinese solar equipment.

Both suppliers are under pressure from China's solar manufacturing glut. Jingsheng's revenue fell about 40 percent year over year in the first half of 2026, and Dalian Linton posted a small loss over the same period, Electrek reported. A US order of that size matters for either.

One detail complicates the framing. Linton's executive board member Todd Barnum called the suit a defence of American innovation and pointed to the company's Rochester roots going back to 1952. But its parent is listed on the Beijing Stock Exchange. As Electrek notes, this is a fight between two Chinese companies, held in a Texas courtroom through American affiliates.

The fleet that is a press release

Further down the credibility ladder, KIDZ AI announced on 30 September that it had sent Tesla a procurement inquiry for 500 Robotaxis. The company, formerly an online coding school called Classover Holdings, is worth about 2.4 million dollars, Electrek reported. Its Q2 2026 results show service revenue down 34 percent year over year to 0.48 million dollars and a 2.5 million dollar quarterly loss. It reported 8.88 million dollars in cash and restricted cash at the end of June.

At Elon Musk's own under 30,000 dollar target for the Cybercab, 500 vehicles would cost roughly 15 million dollars, about six times the company's market value. KIDZ AI's own release says the inquiry does not constitute a binding order and that it may decide not to proceed. The stock rose about 10 percent pre-market and fell about 6 percent after the open.

Electrek places this in a lineage that includes the company's abandoned Solana treasury plan and its later pivot to Hyperliquid, GPU cloud and a Cambodia data centre stake, four hot sectors in roughly 15 months.

There is a real robotics market under all this noise, and it is being built by companies measuring joint angles to the millimetre and shaving milliseconds off control loops. The question the sector has not answered is whether the capital chasing humanoid silhouettes will still be there when the interesting work turns out to be unglamorous: capture rigs, benchmark discipline, and a furnace in Texas that has to pull its first ingot before any of the rest matters.