Tesla's December Optimus reservations, Agility's $465M listing, and Unitree's H2-X sheds a leg on stage

GXO and Agility expansion — GXO is extending its Robots-as-a-Service deal with Agility to two more sites — Statesville, North Carolina and Lebanon, Tennessee — for e-commerce fulfillment. Digit 4 units are slated to arrive in early 2027, moving to Digit 5 once that machine is commercially available. The original Flowery Branch, Georgia deployment, running since the June 2024 agreement, has handled more than 100,000 totes. A first customer coming back for a second and third site is the strongest evidence yet that a humanoid contract is paying for itself rather than serving as a showcase. It also lands days before Agility's Nasdaq debut, which is presumably not a coincidence. agilityrobotics.com

Audi Ingolstadt pilot — Announced at NVIDIA GTC Berlin on October 21, Audi is running a proof of concept with Agile Robots' mobile humanoids at its Ingolstadt Production Lab, picking components out of containers and arranging them in assembly order. Dassault Systèmes and NVIDIA are supplying a virtual twin — DELMIA on the 3DEXPERIENCE platform with Omniverse libraries — to test feasibility, timing and safety before any wider rollout. No throughput figures were published. Parts sequencing is the use case the automakers keep converging on: BMW is doing it with Figure at Spartanburg, and Boston Dynamics is training Atlas for exactly this at Hyundai's Georgia campus. Audi joining makes it four. Audi says its material-supply operations involve about one million picks every day — that's the workflow's scale, not what the robots have demonstrated. humanoidsdaily.com

Reward AI OM-1 — On October 27 Reward AI published a demo of OM-1 doing last-mile delivery (including climbing onto a Cybertruck tailgate for a parcel), picking packages and a sock on stairs, mopping and opening a fridge. The company labels the footage fully autonomous and at normal speed — its own description, not an independent check. Skills come from a shared policy trained on human demonstrations recorded with a wearable hand-tracking rig, plus a low-level controller learned in simulation. The interesting part is the data pipeline rather than the choreography: Reward AI claims a new task needs fewer than 30 minutes of human demonstration, and says it uses no teleoperation or on-robot data for the policy. What the video doesn't show: success rates, how many attempts were filmed, or how the tasks hold up in unfamiliar rooms. humanoidsdaily.com

AGIBOT Jinan dance record — On October 24 in Jinan, AGIBOT ran 110 full-size Expedition A3 humanoids through a synchronized routine lasting five minutes and 43 seconds, with a Guinness World Records judge on stage certifying the count. Coordination ran on the LinkBots fleet system over 5G and Wi-Fi with ultra-wideband positioning, and AGIBOT says the machines were ordinary production units, not hand-tuned for the show. Fleet control at that scale is the real news, not the dancing — it's the same software stack a multi-robot warehouse or retail deployment would lean on. One caveat worth knowing at parties: Guinness already lists 1,372 small UBTECH Alpha 1S robots dancing from a 2018 Rome event, so this record's category is narrower than the headline suggests. humanoidsdaily.com

“the machines would subsequently be deployed in public services, tourism performances, exhibition spaces, research and education”

Agility SPAC close — Agility and Churchill Capital Corp XI said October 30 the business combination should gross roughly $465 million before expenses: about $164 million from the trust after redemptions, a $201 million PIPE at $10 a share led by Foxconn, and $100 million of equity already banked. The deal is set to close November 4, with trading as AGLT starting November 5. This makes Agility the first US-listed pure-play humanoid company with live commercial deployments, and it arrives the same week as the GXO expansion — the customer proof and the capital landing together. agilityrobotics.com

Atlas supply chain — Per KED Global, Boston Dynamics has settled on a Korean-led component network for Atlas and asked suppliers to be ready for purchase orders spanning 10,000 robots — with Hyundai Mobis on all 31 actuators, Samsung SDI and LG Energy Solution on batteries, and NVIDIA supplying the processor. First orders could land as early as late 2027 or early 2028. This is the clearest signal yet of the volume Boston Dynamics and Hyundai are actually designing the Atlas ramp around, and it hands Korean automotive suppliers a concrete route into humanoid manufacturing. Treat the number carefully: KED says quantities and prices are still undecided — this is supplier preparation, not a signed order. KED estimates actuators account for roughly 70% of Atlas component costs. humanoidsdaily.com

Physical Intelligence round — Bloomberg reports the robot-model company is discussing a raise of at least $1 billion at a valuation near $20 billion, with existing backers Avenir and Lux Capital in talks to lead. The round is not finished and none of the parties commented. Bloomberg adds that PI's previous raise valued it at $11.5 billion and that it still has no revenue. The valuation math is the story: from $2.4 billion in late 2024 to a reported $20 billion in under two years, priced entirely on the bet that one shared model can run many robot bodies. The API beta PI opened on October 13 is what investors are reportedly buying. PI's own framing of the hard part comes from co-founder Sergey Levine, on the data rather than the architecture. humanoidsdaily.com

“The most important thing for building foundation models, paradoxically, is not the model itself—it's the data”

Unitree Q3 2026 — In its first earnings report since listing in August, Unitree booked third-quarter revenue of 665 million yuan, up roughly 70% year-on-year, while net profit attributable to shareholders slipped 17.4% to 112.7 million yuan. The reason is a research bill that grew nearly 3.5 times to 127 million yuan — about 19% of the quarter's revenue — going to embodied-intelligence models, compute and headcount, per Reuters. Profit missed the 224 million yuan analyst forecast. Rare hard numbers from the company behind the most-shared videos in the field, and a useful corrective: this is company-wide revenue covering quadrupeds as well as humanoids, and it shows the demos are being funded out of a real but thinning margin. For the first nine months, revenue reached about 1.82 billion yuan, up 55.7%, with net profit of 386.7 million yuan — a figure flattered by a one-time share-based expense in the 2025 comparison. humanoidsdaily.com

Machenta — raised a $40 million seed, Business Insider reports. The London startup is building an AI model that teaches robots a new physical task after seeing it performed just once, betting it can avoid the mountains of real-world data collection its larger rivals are chasing. If the one-demo approach holds on real hardware, it attacks the data bottleneck that currently makes every new robot task expensive. The round was reported October 27; no valuation was given. techmeme.com

Meta and Fourier — The Information reports, citing sources, that Meta spent roughly three months building a humanoid prototype together with Fourier, the Chinese robotics company, and is already at work on a follow-on machine. Meta's only on-record statement is that it has no intention of selling these units. It's Meta's most concrete hardware move in the field, and the shape of it is telling: a fast, externally-built body for internal use rather than a product — which suggests Meta wants a platform to run its AI on more than a robot to ship. The report is unconfirmed; neither Meta nor Fourier has described the machine publicly. techmeme.com

GO-3 Preview — AGIBOT's new embodied foundation model, presented at Innovation Day on October 29, is built around one continuing learning process: pre-training on human first-person video and robot trajectories, then practice via Genie Evolver — a hundred-robot physical fleet blended with physics simulation and world models — and finally learning on the job through Genie Master, which consolidates each day's outcomes into memory and fresh training data. AGIBOT claims first place on three benchmarks, none independently verified, and says the model will be fully open-sourced by the end of 2026. The pitch is robots that improve without new human demonstrations — AGIBOT says a Genie G2 got better across convenience-store tasks this way, calling it recursive self-improvement in the physical world. It hasn't published the evaluation that would let you judge how autonomous that improvement really was. No license, release date or training resources have been specified for the promised open-source drop. humanoidsdaily.com

IRON solid-state battery — XPENG vice president Gu Jie told the SAECCE conference on October 28 that the solid-state battery designed around IRON's tight internal space has begun small-batch output — while making clear the robot itself has not reached start of production. No energy density, runtime, cycle life or volume figures came with the announcement. Runtime is the spec every home-and-store humanoid lives or dies on, so a battery line moving is real news — but the chemistry is left deliberately fuzzy: XPENG's IRON page advertises all-solid-state, while earlier reporting said the production robot would carry semi-solid cells instead because the all-solid-state ones didn't meet install requirements. The assembly lines were commissioned in September with mass production still targeted for late this year and first deliveries next year, so the battery is running ahead of the robot. humanoidsdaily.com

H2-X leg failure — One day after the H2-X launched, the humanoid attempted a side flip on a padded mat at the World Physical AI Conference in Hangzhou last Thursday, wobbled on landing, and its right lower leg came apart at the joint. Attendee footage circulated widely, and TechNode reports Unitree has offered no public explanation, investigation findings or fix plan. This is the machine sold for security patrols and emergency rescue, and a structural joint failure under landing load is exactly the failure mode that matters for those jobs — more than the car-door kicks and four-metre backflips in the launch films. TechNode's analysis points to landing forces of several thousand newtons and carbon-fiber legs with limited structural redundancy once loads exceed design limits. technode.com