Robots digest — Sep 23, 2026
Top line: The robotics developer-platform layer consolidated under silicon — Qualcomm is buying PickNik (MoveIt stays open-source) and wiring it into Dragonwing + Arduino — while Wayve's CEO put numbers on the humanoid gap (20–50% task success, "where driving was five years ago") and China's state-led buildout kept grinding without much that's new.
Qualcomm acquires PickNik — the de facto manipulation stack joins a chip vendor (new)
What happened: Qualcomm Technologies agreed to acquire PickNik Robotics (Boulder, founded 2015), stewards of MoveIt — the open-source ROS manipulation framework most of the field actually uses. Qualcomm says MoveIt 1 and 2 stay open-source under existing licenses with community-driven roadmaps, hardware-agnostic support across third-party vendors continues, and both companies are OSRA founding members (Space ROS also continues). Integration starts with Arduino VENTUNO Q boards (Qualcomm bought Arduino in Oct 2025) and Dragonwing platforms; MoveIt Pro on Dragonwing is being demoed at ROSCon Toronto this week. Deal amount unspecified, subject to closing (Robot Report, ~7:29 AM ET). Same window, same company: a Robot Report piece on the VENTUNO Q board as Arduino's "first intentional step" into robotics (~2:04 PM ET).
What's genuinely new: the manipulation-middleware standard is now inside a chip vendor's platform — the third platform-consolidation move in days (Intrinsic open-sourced Sep 22; NVIDIA-Hugging Face this month). The stated economics are Qualcomm's: silicon + boards + middleware sold as one path "from prototype to production." The open questions are the ones the Intrinsic thread asked: does the hardware-agnostic/open-source commitment survive integration, and does this build Android-grade developer gravity or just a Qualcomm funnel.
Sources therobotreport.com therobotreport.com
Wayve's Kendall: cars first, humanoids "still at the demo stage" — with numbers (new — reality-check datapoint)
What happened: Wayve CEO Alex Kendall at HumanX Amsterdam (~4:21 PM ET, day after the Mercedes-Benz definitive production agreement): Wayve's AI will be in Mercedes consumer cars within two years (third top-10 carmaker in nine months, after Nissan and Stellantis), a Tokyo robotaxi service by year-end (following supervised, safety-driver London operations with Uber), and — on humanoids — they are "still at the demo stage," working 20–50% of the time on custom prototype hardware, which he says is "where driving was five years ago." Fewer than 10,000 robotaxis exist worldwide vs ~100M cars sold a year; Huang's "physical AI ChatGPT moment" is "not yet" because a hundred million people don't use physical AI.
What's genuinely new: the AV side of physical AI joining the humility chorus — with quantified claims instead of vibes, from a CEO who has real production contracts to back his position. Both halves are ledgerable: Tokyo by year-end, Mercedes in cars by ~2028. The 20–50% success figure is a concrete external benchmark for the reality-check war (vs Agility's filed $1.8M revenue and Hyundai's "not yet at scale").
Sources thenextweb.com/news electrek.co/2026
China cluster: a week-late "robots building robots," elder-care deployments, IPO curbs (continuation — china-industrial-shift)
What happened: Khaama's "China opens smart factory where humanoid robots build other robots" (~4:25 AM ET) is a week-late recirculation of UBTECH's Liuzhou factory opening (production started Sep 12; company claims: one humanoid every 10 minutes, 10,000/yr target — Global Times Sep 14). DW ran a video piece on humanoids moving into elder care (deployment color for the aging-population demand case). And a Portuguese-language syndication re-ran the humanoid IPO curb record (regulators braking listings as hype outpaces demand).
What's genuinely new: almost nothing — diffusion of items already on record. The only quantified adds (10-min takt, 10,000/yr target) are UBTECH's own claims from Sep 14, company-reported and unaudited. The thread's standing frame holds: production claims keep circulating without unit counts that survive audit.
Sources khaama.com globaltimes.cn/page dw.com/en
Polybot's tomato harvester learns from people, then picks full rows (new, small)
What happened: Polybot — a Tübingen startup spun out of the ELLIS Institute and Max Planck Institute for Intelligent Systems — has an autonomous tomato-harvesting robot that learns picking from human demonstrators and is now picking full greenhouse rows; deliveries to growers could begin early 2027 (surfaced via MSN syndication of The Cool Down/Interesting Engineering).
What's genuinely new: agricultural imitation learning outside the humanoid circus — narrow task, real greenhouse, commercial date. Fits the deployment-not-demos frame; the early-2027 delivery date is the checkable claim.
Sources interestingengineering.com/videos thecooldown.com/green-tech
One-liners: Agility CTO Pras Velagapudi on robot data security once humanoids leave controlled environments — "faces, address labels, proprietary IP" — the aftercare/governance layer getting CTO airtime (~12:28 PM ET). Boston Dynamics posted a Stretch endurance-testing engineer interview (~10:31 AM ET) — the wear-and-maintenance layer continuing its disclosure pattern.
Dropped: Toyota's $6.4B/400k-robots recirculation (the dollar conversion was printed Sep 20; no new facts); Samsung's "world-first team surgery" recirculation (the Sep 16 staged demo, covered Sep 17 — the headline re-oversells what the snippet itself calls a simulated OR); Vogue World Milano runway robots, AheadForm's bionic robot at a Shanghai conference, the Beijing hurdle stumble, cage-combat teleop and the LaPenna "human vs robot fight" (spectacle layer); battery-free textile sensors (off-beat research); Light-O1's embodied foundation-model announcement (single-syndication claim, no deployment evidence); the RoboBusiness preview cluster (State of Humanoids, Intel, outdoor robots — trade-show promos); content-mill explainers ("What Is Robotics Transforming…", the note.com physical-world-simulation piece).