Three Days, 21 Robot Stories: Orders Are Real, Prices Are Falling, Touch Is Entering the Policy

By the Latest Robot News editorial team. Compiled from Jijia Juexing’s daily industry brief covering 23–25 October 2026.

Twenty-one items in seventy-two hours reads, at first glance, like another funding-and-launch cycle. Sorted by category — orders, capacity, incidents, valuations — three lines actually moved: humanoid supply-chain orders turned into countable units, mass production began to be tested by real takt time, and the primary market shifted from grabbing allocation to picking winners.

1. The supply chain: order numbers became hard data

The most useful numbers of the three days came from Morgan Stanley’s CIIF survey note of 19 October:

The same note flags price pressure: Leaderdrive’s ASP is still RMB 800–900 while competitors quote RMB 300–400. For the first time in this cycle, humanoid reducers are an order book counted in units — and for the first time, a price war is visible on the horizon.

The implication for domestic joint and reducer makers is narrow and uncomfortable: volume proves demand is real, ASP erosion proves supply is arriving fast. The next contest is not whether you can make the part, but whether one line can hold both consistency and cost.

2. Mass production: from “rolled off the line” to “ramping”

Tesla confirmed in its Q3 2026 results that Optimus has entered production at Fremont, moving from line construction to an early ramp, with a second line in Texas still under construction. The company describes the first units as meeting production-grade quality requirements, with hardware design and manufacturing processes still iterating, but the existing process proven suitable for commercial production. The Fremont line was previously planned for eventual capacity of one million units a year.

On the earnings call, Musk added the training side: Tesla is building a facility called Optimus Academy that will put the first 10,000 to 20,000 Optimus units to work on autonomous practice and a broad set of tasks, feeding the collected data back into the model — compensating for the fact that Optimus, unlike FSD, has no global fleet generating training data. At USD 50,000–100,000 per unit plus other costs, the cluster represents roughly RMB 14 billion of investment.

Qiyuan Robotics, under SWANCOR Advanced Materials (上纬新材), supplied the consumer-grade reference point: its 5,000th personal robot came off the line on 25 October, one month after the Q1 and T1 launched on 20 September. The line now produces one unit every 2.5 minutes, with daily capacity up from about ten to more than 200. Some delivered units show more than one hour of daily use and more than 150 effective interactions per day.

A one-million-unit plan entering an early ramp, and one unit every 2.5 minutes in practice. Read together, “mass production” in 2026 has tiers: getting a unit off the line is the entry ticket; takt time and yield are the score.

3. The incident: Unitree H2-X lost balance mid-flip

On 22 October at the 2026 World Physical AI Conference, a white humanoid performing a side flip landed off balance and fell, with one lower leg separating at the connection point. The robot was the H2-X Unitree had launched the day before, marketed for security and emergency response with an official video showing it denting a car door, lifting a car, and flipping from a platform about four metres high, with up to eight hours of endurance. Unitree had not commented at the time of reporting.

Public failures of high-dynamics demos are not rare. The question they raise is practical: a public demo of a dynamic manoeuvre tests more than the controller — it tests structural connections, fastening and impact design. For component suppliers the lesson is direct. Between demo-grade and delivery-grade reliability, the gap is usually mechanical detail.

4. Models and data: touch enters the training loop

AgiBot + Peking University released ViGAR (visual-goal-conditioned action reasoning), described as the first complete world-action model for full-size bipedal humanoids, validated on multi-step manipulation on the AgiBot A2 platform. On the RoboTwin 2.0 Clean2Random benchmark it reports 74.51% average success, 12.81 points above the strongest baseline; on five long-horizon real-robot tasks, average completion 0.69 versus 0.60 for π0.5. Paper, evaluation code and weights are public.

Fudan University’s RoboSkill lets a generalist robot agent save effective on-site methods together with the code it actually executed as reusable skills, with tactile feedback accelerating confirmation of outcomes. With touch added, GPT-6-Astra’s simulated average time fell from 25.35 to 15.93 minutes; skill reuse cut GPT-5.6-Sol from 94.5 to 35.9 minutes; and when skills went from text-only to text-plus-code, first-attempt success rose from 15% to 67.5%.

NVIDIA expanded Isaac at GTC Berlin: a reference humanoid from Agile Robots (Agile ONE SN), the Isaac OS runtime, Isaac Lab 3.0, an open reference workflow for DGX Station, a SimReady central library with more than 2,000 simulation assets, and the Human 1M dataset. Agile ONE SN runs Jetson Thor with Isaac GR00T preconfigured, stands 1.4 m with 31 degrees of freedom and two 20-DoF five-finger hands, and is planned for early 2027.

Unitree’s Wang Xingxing said the company is pre-researching a “Physical AI robot self-evolution V1.0” model that would let AI run the full loop of literature search, code writing, simulation training and result evaluation, reducing dependence on manual programming and tuning. He expects physical AI to break through within one to three years, and argues the core obstacle to embodied AI’s ChatGPT moment remains the alignment error between model input/output and the real robot.

5. Capital: heat intact, expectations turning cautious

• Eka Robotics announced a USD 150 million Series A on 20 October, led by a16z and F-Prime with Samsung Next participating, at a reported post-money valuation of USD 1.45 billion. Founded by MIT professor Pulkit Agrawal and former DeepMind researcher Tuomas Haarnoja, it pursues a vision-force-action (VFA) route rather than mainstream VLA, trained with large-scale simulation reinforcement learning; its robots grasp up to 1,500 items per hour, about 2.5x the human baseline, and are in pilot with industrial customers.
• Fujian Humanoid Robot Innovation Center closed a Pre-A+ round with Fuzhou Gulou district construction investment group and an Xiamen incubation platform, after an earlier round from the Minjiang industry guidance fund. Founded in September 2025, its FUZHI line runs a self-developed f-VLA model and a fully 3D-printed lightweight body that pushes system cost below RMB 10,000 — about 5–10% of comparable products — with a Minqing plant planned for 10,000 units a year.
• Higher IPO thresholds, diverging views. After regulators signalled tighter listing standards for robotics companies, market reports said Songyan Dynamics had confidentially filed for a Hong Kong IPO seeking several hundred million dollars, and that AgiBot had opened a RMB 1–1.5 billion round at a reported RMB 50 billion valuation. Investors are splitting: Cornerstone Capital says it will still back companies with genuine differentiation, while one state-backed investor said it has essentially stopped reviewing embodied AI deals since June, citing heavy homogenisation and a lack of substantive technical progress.

6. Deployment: three scenes that do not look like robots

Pharmaceutical distribution. Kangbaijia’s logistics centre in Fuzhou now runs hundreds of box-climbing robots carrying medicine crates through racks more than ten metres high, with sorting robots dropping items into the right chutes. The centre is the master warehouse for more than 2,000 stores across Fujian, Zhejiang and Jiangxi; stores in Minhou can be restocked in as little as 30 minutes.

Commercial cooking. Data from the China Hotel Association puts cooking-robot penetration in Chinese fast-food chains and mass-market restaurants at 30%. Two robots in a Shenzhen chain outlet produce 600–700 dishes a day; three units at a Guangzhou school canteen, each handling nearly 20 kg per cook cycle, complete a thousand meals in three hours. The World Cooking Robot Industry Development Report projects the Chinese market will exceed RMB 11.7 billion by 2030.

Bridge fabrication. At China Railway Shanhaiguan Bridge Group’s Nantong smart factory, white welding robots work inside the cramped cells of the Zhangjinggao Yangtze River Bridge deck. The design — a world-first longitudinally open L-rib with apple-shaped holes in an orthotropic steel deck — leaves gaps too tight for manual welding to guarantee quality. Portable and rail-mounted collaborative robots scan and weld without teaching; after the upgrade, overall welding efficiency rose 20%, daily output from 10 to 20 tonnes, pass rate up 10%, with about 700 tonnes of qualified members produced.

Six more worth noting

• Galbot + Unitree announced joint development of Galbot ET1, described as the first natively agentic robot, first shown at the World Robot Conference in August.
• Hikrobot launched its “Mingzhi” single-unit intelligent products and the industry’s first cross-aisle STU tote robot, with deployment from unboxing to production in as little as 30 minutes; it held the global top share in mobile robots in 2025 and passed 200,000 cumulative units across categories by July 2026.
• Cerence AI × Geely extended their partnership from smart cockpits to embodied robots, with multilingual speech recognition covering French, English and Mandarin, shown at the Paris Motor Show in October; more than 525 million vehicles use Cerence technology.
• Lushan New Materials showed a fully flexible stretchable electronic skin with a 100 mg detection floor, sub-millimetre spatial resolution, 1,000 Hz refresh across a hundred-array, and cycle life beyond hundreds of thousands of cycles, with proximity sensing and conformal attachment.
• Camera robots. Miaodong Technology’s Beni terrain-following robot is due this month — two-wheeled, active following, 4K capture, from USD 549 on Kickstarter; Xingshi Technology closed a million-dollar strategic round; Insta360 published its Cameraman vision in July.
• Shangpin Home + Qiyuan opened an 800 m² experience centre in Guangzhou’s financial district; the joint venture Shangyuan Zhiqi also open-sourced the WorldSimReady-Home simulation dataset, downloaded 27,000 times on Hugging Face in its first month and topping the Robotics trending list.

Three conclusions from the three days

The orders are real, and so is the price erosion. Reducer makers are booking units, not framework agreements — while ASPs face quotes at RMB 300–400 against a current RMB 800–900. Differentiation in 2027 will not be technical; it will be structural cost.

“Mass production” now needs tiers. A one-million-unit plan is not the same claim as one unit every 2.5 minutes, and the H2-X fall is a reminder that demo-grade and delivery-grade sit on opposite sides of structural and connection design.

Touch is entering the training loop. The Fudan numbers are the most persuasive of the three days: skills going from text to text-plus-code lifted first-attempt success from 15% to 67.5%, and adding touch cut task time from 25.35 to 15.93 minutes. Sensing is no longer only a hardware spec — it is a variable in model efficiency.

Source: Jijia Juexing daily industry brief, 23–25 October 2026; underlying items from IT Home, China Electronics News, NVIDIA, Securities Times, Star Market Daily, CNR, China Economic Net, Jiemian News, QbitAI, 10jqka, Fuzhou Daily, CCTV, Xinhua Daily, Shanghai Securities News, DeepTech and Sina Finance. Figures as reported by those sources and not independently verified by Latest Robot News.