Optimus starts rolling off the Fremont line; AGCO signs Digit for Kansas; Unitree films kung fu and warns its own investors

Tesla Optimus — Tesla's Q3 2026 shareholder deck and October 21 earnings call: robots the company calls production-quality are now being built on the Fremont line, which the deck classifies as an early ramp while the Texas factory is still construction. Elon Musk called it an early start, put external deliveries in the late-2027 window, and said Fremont already holds most of the equipment for a one-million-robot-a-year line. This is the first concrete manufacturing status plus a delivery timetable from the program, and it keeps the hands problem in view: Tesla is building robots with hand designs that are still unfinished. Musk wants 10,000–20,000 robots practicing tasks through self-play for training before outside sales; the AI5 chip is targeted for roughly April 2027. Humanoids Daily

“The bots coming off this line are the first production-quality bots that we have manufactured”

XPENG IRON — XPENG posted a clip of IRON following a person's hand as the offered greeting changes — handshake, then fist bump, then high-five — and switching to its other arm when the target drifts out of reach, with no spoken commands. The company credits VLA 2.0 and says the robot refreshes its read of the scene every 0.2 seconds, about five updates per second. The footage carries an R&D test-build label; no success rate or end-to-end latency is given. A small but genuine interaction skill — adapting mid-gesture rather than completing a fixed action — and the company is honest about the label. XPENG doesn't say whether the VLA 2.0 build in the clip is the same model its cars run. Humanoids Daily

“Natural interaction begins with these simple moments.”

Unitree H2-X — Unitree's October 21 film, Kung Fu Legend Guardian, shows the H2-X striking a car door hard enough to break glass, backflipping off the roof and leaping from a platform marked four meters, then staged guardian scenes: stepping in against a knife-wielding actor and lifting a car to free a trapped dummy. Unitree says the footage is real and unsped. The advertised ~8 hours of endurance is on-screen defined as standing endurance, and the scripted encounters are promotional narrative, not evidence of autonomous threat detection. Watch it for the acrobatics, not the rescue claims — the company itself is simultaneously telling investors commercialization is early. Unitree says the robot can be run over by a vehicle in the film and get back up; the film does not explain how much human direction the sequences took. Humanoids Daily

Tuesday Labs Voxel — Tuesday Labs, founded by former Boston Dynamics staff roboticist Ros Shinkle, unveiled Voxel, a legged home robot pitched around four chores: resetting rooms, finding misplaced items, fetching them and tending plants. It takes $200 refundable reservations for the Series-1 against an estimated $20,000 price, with US deliveries expected from 2028. The reservation agreement says the robot shown is a development prototype headed for a substantial hardware redesign. A new entrant in the home-robot race 1X is running, and interestingly it bets on simple one-degree-of-freedom grippers rather than multi-fingered hands. Humanoids Daily

Eka Robotics — Eka, founded by MIT professor Pulkit Agrawal and former DeepMind researcher Tuomas Haarnoja, closed a $150 million Series A led by a16z and F-Prime at a reported $1.45 billion post-money valuation. Its bet is a vision-force-action model — not vision-language-action — plus a force-sensing gripper, trained by reinforcement learning in simulation rather than on human video. The claim veteran roboticists found striking is speed with contact: about 1,500 picks per hour of objects up to 1 kg, versus roughly 600 for a human, at over 95% success — numbers to watch for in field pilots, since lab-to-line is the unproven step. The gripper hardware costs under $1,000 and is rated for 5 million duty cycles; the target is 20 million, the bar set by standard industrial arms. TechCrunch

“Vision tells you roughly where the object is, what the pose is, but the actual grabbing, the actual searching is happening through forces.”

Noetix Robotics — Bloomberg says it has learned that Noetix, the Beijing maker of the Bumi, N2 and E1 humanoids, has quietly submitted listing paperwork in Hong Kong that might raise a few hundred million dollars, with a debut possible during 2027 at the earliest. Treat it as unconfirmed: Bloomberg's sources say deliberations continue and the deal's size and timing could still shift, and Noetix declined to comment. A second Chinese humanoid maker testing the public markets right after Unitree — whose shares have slid nearly 40% from their post-IPO peak — tells you investors are already pricing in adoption risk. The company banked about 1 billion yuan, roughly $150 million, in March, in a round led by an investment platform tied to battery giant CATL. Humanoids Daily, citing Bloomberg

YITA — It builds force-controlled arms, integrated actuators and joint modules for embodied-AI developers — the layer many humanoid makers buy rather than build. Humanoid makers themselves investing in an actuator supplier is a supply-chain signal worth more than the modest sum. Its products include compliant dual-arm systems and manipulation platforms that regulate contact forces while handling objects. TechNode

Enact Intelligence — Zurich's Enact has come out of stealth with a $30 million pre-seed from redalpine, Notion Capital and firstminute — reported as the largest pre-seed ever raised in Switzerland — with Yann LeCun and Mistral's Arthur Mensch among the angels. Its idea: read touch from the motors rather than fragile external skin sensors, the way people sense tension in muscles, and target small-batch factory assembly like cable insertion and seal seating. Founder Manuel Yves Galliker previously led the software that taught 1X's NEO to walk, so this is a credible bet that dexterity is bottlenecked on cheap, robust force data. In an NZZ workshop interview, Galliker set a goal of sampling force above one kilohertz from each motor and predicted training for a new task could compress from months to days — stated aims, not published results. The Next Web

“We can build on top of existing models, like video models, and adapt them to output robot actions for interacting with the real world”

Figure 04 — Figure made public a recording of a staff meeting on October 22 in which Brett Adcock told employees the fourth-generation robot's design is frozen — its actuators already running in the test lab and ten full battery packs powered up the week before — and sized the AI push around it: 37.5 million videos uploaded to the Index human-video program across 112 countries, roughly $8.1 million spent on data in a single week, and over $2 billion in the bank. The number that matters is the burn rate on data: Figure is spending like a frontier lab and betting the whole roadmap on human-video pre-training for Helix 3.0. Humanoids Daily

“He characterized the internal result as “baby AGI,” but said it was not ready to share in full.”

Physical Reasoning — AGIBOT's fourth AGIBOT WORLD release, announced October 22, pairs robot action trajectories with question-and-answer annotations covering nine reasoning skills — search, quantity judgments, memory, subtask ordering among them — so a model learns to pick a different action path when starting conditions change, as in restocking a minibar that must end with three full bottles. The release includes both successful and erroneous runs, annotated for whether the right object was chosen. This is the sector moving from hard movements to decisions, and AGIBOT says it is already training its next foundation model on the data. Note the license: CC BY-NC-SA 4.0, so research use, not commercial. No episode count for Release 4 and no measured task-success improvements are given in the announcement. Humanoids Daily

Schaeffler Taicang — At a new plant of roughly 7,500 square meters in Taicang, China, six humanoids working five stations now assemble the linear and rotary actuators that go into other humanoids' joints, in what Schaeffler calls the start of mass production for those components. The site doubles as a training and data-collection facility with a digital twin of the line, letting changes be tested virtually first. A supplier running humanoids on its own actuator line is the clearest picture yet of the robot-builds-robots loop — and the order number says demand is real, at least as bookings. Schaeffler counts its global humanoid-robotics order book at around €400 million, helped by two new Greater China customers in Q3; that figure is incoming orders, not recognized revenue, and no production rate or intervention figures were disclosed. Humanoids Daily

NEO's 25-DoF hands — detailed the tendon-driven hands it says ship on every NEO: 25 degrees of freedom — 22 fully actuated in the fingers and palm, 3 at the wrist — all force-controlled and backdrivable through quasi-direct-drive tendons at gear ratios of roughly 5:1 to 15:1, so contact forces flow back through the motors instead of dying in a gearbox. The hands are IP68 waterproof and food-safe. The low gear ratios are the engineering point: every joint doubles as a sensor, which is what makes learning from every grasp possible. Peak torque reaches 3.5 Nm at the thumb CMC, with distal flexion forces up to 45 N and ±0.2 mm positioning accuracy. 1X says a dedicated in-house line gives it capacity to produce 10,000 hands this year. 1X

Announcement 2026-003 — Three sessions of gains big enough to trip the STAR Market's abnormal-trading threshold — closing prices deviating more than 30% from the benchmark across October 16, 19 and 20, after a 16.15% jump to ¥495.92 — forced Unitree to file a formal notice on October 21. In it, the company warns that humanoid commercialization is still early-stage, that slower industry adoption could hurt its revenue and earnings, and that investors face the risk of a valuation snap-back. The same company releasing backflip films is formally cautioning the market, and the valuation gap it cited is the quotable part. Humanoids Daily