ACQUIRING TARGETS...
· DroidHunt news
Generative Bionics, the Genoa-based maker of the GENE.01 industrial humanoid, announced at CES Unveiled Milan (Oct 21) a proprietary three-site "Physical AI Data Factory" network: a first 5,000-m² hub in Italy targeted to be operational by Q4 2027, then a second site elsewhere in Europe and a third for global markets. Fleets of GENE.01 robots will work in reproduced industrial environments, feeding a real-to-sim-to-real loop — data captured physically, stress-tested across thousands of digital-twin variations on AMD Helios rack-scale systems, then validated back on robots. AMD, an investor in the company's EUR 70M seed round, is the compute partner. The plan extends the company's Eni evaluation pact (DroidHunt, Oct 2).
Why it matters: The announcement frames training data — not hardware — as Europe's humanoid bottleneck, and bets on sovereign data infrastructure as the fix. It is also a concrete landing spot for AMD compute in robotics, parallel to NVIDIA's Jetson Thor ecosystem.
Sources: prnewswire.com therobotreport.com
Teradyne Robotics — parent of Universal Robots (UR) and Mobile Industrial Robots — and Stuttgart-based Sereact announced an official partnership (Berlin, Oct 21; The Robot Report, Oct 25): Sereact's Cortex vision-language-action model runs on UR cobot arms to automate unstructured retail and logistics tasks. The companies say hundreds of UR robots in dual-arm systems have shipped to live European customer sites, autonomously processing returned fashion items for Zalando with CEVA Logistics at fulfillment centers in Greven, Germany and Świebodzin, Poland, with US expansion planned via Sereact's new Boston HQ. Cortex runs as one production model across returns handling, piece-picking, kitting and machine tending (shipment counts company-claimed).
Why it matters: This is the brain layer monetizing without humanoid bodies: a VLA running on fixed dual-arm cobots, at claimed fleet scale, in production. It is direct competitive context for the humanoid economics thesis — and shows Jetson Thor-class compute spreading into installed industrial arms.
Sources: finance.yahoo.com therobotreport.com
Tags: teradyne-robotics deployment europe adjacent confirmed
An arXiv preprint (2610.31859, posted Oct 22) finds that when a vision-language-action model is fine-tuned in bf16 with no fp32 master weights, AdamW rounds away any update smaller than 1/512 of a weight — silently erasing at least 79% of updates in LIBERO test runs. The authors found 9 of 12 open VLA codebases, including LeRobot's SmolVLA defaults and Physical Intelligence's π0/π0.5 guides, train this way; 99% of sampled full-weight SmolVLA fine-tunes on Hugging Face Hub store the action expert in bf16. Keeping trainable weights in fp32 raised LIBERO-Object success from 4.3% to 49.3% for SmolVLA and from 5.1% to 80.9% for π0.5 — a one-line fix. Peer review pending.
Why it matters: If the finding holds, a large share of published VLA fine-tuning results may be undertrained by accident — a cheap fix with outsized effects on reproducibility and on anyone building products on fine-tuned open VLAs.
Sources: arxiv.org
Tags: hugging-face research us reported
← Newer: Oct 27, 2026 Older: Oct 23, 2026 →
Home Every digest Robot catalog Drones on DronePriceHunt Privacy & affiliate disclosure RSS