<?xml version="1.0" encoding="UTF-8"?><rss version="2.0" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Robotics Daily · Brains and Parts</title><description>Every Robotics Daily capsule in the segment Brains and Parts, the latest first.</description><link>https://roboticsdaily.robotshop.com/</link><language>en</language><atom:link href="https://roboticsdaily.robotshop.com/feeds/brains-and-parts.xml" rel="self" type="application/rss+xml"/><item><title>Runway releases Praxis-1, an open-weight model that drives robots</title><link>https://roboticsdaily.robotshop.com/capsules/runway-praxis-1-open-weights-2026-11/</link><guid isPermaLink="true">https://roboticsdaily.robotshop.com/capsules/runway-praxis-1-open-weights-2026-11/</guid><description>Runway released Praxis-1 on 11 Nov. 2026, a 27-billion-parameter world action model for robots, with weights and fine-tuning code under the Apache 2.0 licence.</description><pubDate>Mon, 16 Nov 2026 00:00:00 GMT</pubDate><content:encoded>&lt;figure&gt;&lt;img src=&quot;https://roboticsdaily.robotshop.com/_astro/runway-praxis-1-open-weights-2026-11.BnY-b58W_Z2jxaxa.jpeg&quot; alt=&quot;A robot arm with a black multi-camera gripper reaches toward a red foam block on a shelf beside a vintage camera, with &amp;quot;Introducing Praxis-1&amp;quot; and the Runway Robotics wordmark overlaid on the frame.&quot; /&gt;&lt;figcaption&gt;Image: Runway / Praxis-1 announcement · press-kit&lt;/figcaption&gt;&lt;/figure&gt;&lt;p&gt;Runway, the AI video company, released Praxis-1 on 11 November 2026. It is a world action model: a 27-billion-parameter mixture-of-experts diffusion transformer that predicts video and robot actions together, so the same network imagines what happens next and decides how to move.&lt;/p&gt;&lt;p&gt;The weights, inference code and fine-tuning code are published under the Apache 2.0 licence, which allows commercial use, on GitHub and Hugging Face, with native support for Hugging Face&amp;#39;s LeRobot framework, full fine-tuning and LoRA. Hosted inference is also offered.&lt;/p&gt;&lt;p&gt;On Runway&amp;#39;s own RoboLab-120 benchmark of 1,200 episodes, Praxis-1 reached 46.2% success, against 42.9% for FLUX 3 Action, 36.8% for Cosmos3 Nano and 28.0% for pi0.5. After fine-tuning on Google&amp;#39;s RT-1 data it reached 80.0% on SimplerEnv, against 52.4% for the original RT-1 policy. The results have not been replicated independently.&lt;/p&gt;&lt;p&gt;Launch partners Noble Machines, Standard Bots and Ultra tested it on a legged humanoid, a six-axis industrial arm and a dual-arm warehouse robot.&lt;/p&gt;</content:encoded><category>Brains and Parts</category></item><item><title>Berkeley, NVIDIA and Bosch release Graph-as-Policy for robot workcells</title><link>https://roboticsdaily.robotshop.com/capsules/berkeley-nvidia-graph-as-policy-corl-2026-11-11/</link><guid isPermaLink="true">https://roboticsdaily.robotshop.com/capsules/berkeley-nvidia-graph-as-policy-corl-2026-11-11/</guid><description>The open-source CoRL 2026 Spotlight system has AI agents compose robot skills into a checked graph, reaching 0.93 to 0.99 success where vision-language-action baselines fall to about 0.20.</description><pubDate>Thu, 12 Nov 2026 00:00:00 GMT</pubDate><content:encoded>&lt;figure&gt;&lt;img src=&quot;https://roboticsdaily.robotshop.com/_astro/berkeley-nvidia-graph-as-policy-corl-2026-11-11.Os5o9IXd_23b6fe.jpeg&quot; alt=&quot;A white Franka robot arm in a black-curtained lab workcell drops a small pink object into a woven basket, with a mustard bottle, a can and several boxed items on the wooden table in front of it.&quot; /&gt;&lt;figcaption&gt;Image: UC Berkeley / NVIDIA / Bosch, Graph-as-Policy project page · press-kit&lt;/figcaption&gt;&lt;/figure&gt;&lt;p&gt;Researchers from UC Berkeley, NVIDIA and Bosch, with Ken Goldberg among the senior authors, released Graph-as-Policy (GaP), a CoRL 2026 Spotlight paper, together with its code and a 51-skill library. It targets tasks where a robot repeatedly handles objects of varying shape and position inside a fixed workcell.&lt;/p&gt;&lt;p&gt;An orchestration agent splits a task into segments, and skill agents assemble sub-graphs from the library into one type-checked computation graph. The graph is rehearsed and repaired in simulation before it runs on a robot, with no task-specific training data.&lt;/p&gt;&lt;p&gt;On eight benchmarks, four simulated and four real, the authors report 0.93 to 0.99 success under pose and geometry variation, where vision-language-action baselines fall to about 0.20. A Make Popcorn task rose from 33% to 94% in simulation and to 90%, 18 of 20 trials, on the real robot. On bimanual crate washing GaP scored 0.953 against 0.987 for a hand-engineered graph.&lt;/p&gt;&lt;p&gt;The results are the authors&amp;#39; own, and the project page describes quasi-static workcell tasks, not open-ended manipulation.&lt;/p&gt;</content:encoded><category>Brains and Parts</category></item><item><title>South Korea adds sensors, actuators and batteries to its physical-AI plan</title><link>https://roboticsdaily.robotshop.com/capsules/south-korea-physical-ai-parts-batteries-plan-2026-11/</link><guid isPermaLink="true">https://roboticsdaily.robotshop.com/capsules/south-korea-physical-ai-parts-batteries-plan-2026-11/</guid><description>The Korean government added three core physical-AI component tasks to its Hyper-Innovation Economy plan on 11 November 2026, targeting humanoid batteries by 2030, Seoul Economic Daily reports.</description><pubDate>Wed, 11 Nov 2026 00:00:00 GMT</pubDate><content:encoded>&lt;figure&gt;&lt;img src=&quot;https://roboticsdaily.robotshop.com/_astro/south-korea-physical-ai-parts-batteries-plan-2026-11.TakqCc4j_G0HRJ.jpeg&quot; alt=&quot;KAIST&amp;#39;s HUBO humanoid robot, a Korean-built robot with jointed metal limbs, stands on an elevated metal platform on an obstacle course at the 2015 DARPA Robotics Challenge finals in California, with a worker in a red vest and white helmet beside it.&quot; /&gt;&lt;figcaption&gt;Image: Master Chief Petty Officer John Williams / U.S. Navy, via Wikimedia Commons (public domain) · public-domain&lt;/figcaption&gt;&lt;/figure&gt;&lt;p&gt;South Korea added three core physical-AI component tasks, covering sensors, actuators and batteries, to its Hyper-Innovation Economy Leading Project plan, Seoul Economic Daily reported on 11 November 2026. The plan was presented at an emergency meeting of economic ministers in Seoul and joins 15 projects added in July.&lt;/p&gt;&lt;p&gt;For sensors, the state is putting 49.8 billion won into on-sensor AI chip development between 2027 and 2030. Domestic mass production of actuators is targeted for 2029. For batteries, the plan covers higher output and energy density for existing cells and all-solid-state batteries, with humanoid batteries to be commercialized in earnest by 2030.&lt;/p&gt;&lt;p&gt;The newspaper says the budget proposal for next year includes 13.3 billion won in state funds for 18 tasks across actuators and batteries, and that the plan aims to localize 80% of core robot parts by 2030. The report names no ministry and no primary government release was found, so the figures rest on the newspaper. Other outlets quoted a $330 million total; that figure could not be traced in this report.&lt;/p&gt;</content:encoded><category>Brains and Parts</category></item><item><title>NVIDIA releases GR00T 2 and Isaac Lab 3.0 at CoRL</title><link>https://roboticsdaily.robotshop.com/capsules/nvidia-gr00t-2-isaac-lab-3-corl-2026-11/</link><guid isPermaLink="true">https://roboticsdaily.robotshop.com/capsules/nvidia-gr00t-2-isaac-lab-3-corl-2026-11/</guid><description>On 9 November 2026 NVIDIA said GR00T 2 averages 71% success across 50 robot tasks from one demonstration each, and learns from human video.</description><pubDate>Tue, 10 Nov 2026 00:00:00 GMT</pubDate><content:encoded>&lt;figure&gt;&lt;img src=&quot;https://roboticsdaily.robotshop.com/_astro/nvidia-gr00t-2-isaac-lab-3-corl-2026-11.DaYnc2pF_Z1zAlKH.jpeg&quot; alt=&quot;A four-panel collage of robot simulation and lab scenes: a humanoid robot pulling a drawer out of a washing machine, a robot arm operating a toy crane, a field of dual-arm robots at wooden tables in a simulated training environment, and a white humanoid robot handling lab glassware.&quot; /&gt;&lt;figcaption&gt;Image: NVIDIA / NVIDIA Blog · press-kit&lt;/figcaption&gt;&lt;/figure&gt;&lt;p&gt;NVIDIA introduced GR00T 2 and Isaac Lab 3.0 on 9 November 2026 at the Conference on Robot Learning. GR00T 2 is an open world action model built on NVIDIA Cosmos and trained on robot data and human data, including 100,000 hours of egocentric human video. According to NVIDIA, it completed 40 of 50 evaluated tasks at least once with no task-specific demonstrations, averaged 71% success across the 50 tasks with one demonstration per task, and reached 95% with five. NVIDIA also says it transfers skills learned from human demonstrations between different robot bodies without robot demonstrations.&lt;/p&gt;&lt;p&gt;Isaac Lab 3.0 is the new release of NVIDIA&amp;#39;s open-source robot-learning framework, published on GitHub under the tag v3.0.0. It lets users switch between the PhysX and Newton physics solvers, adds a reinforcement-learning post-training library for foundation models such as GR00T, and models deformable objects like cables and cloth. On a vision-based dexterous manipulation benchmark NVIDIA reports up to 2.5 times the training throughput of Isaac Lab 2.3 with PhysX.&lt;/p&gt;&lt;p&gt;Early testers named by NVIDIA include Nexuni, Foxconn, Siemens and MicroAGI. NVIDIA&amp;#39;s blog calls GR00T 2 open but does not state its licence terms. All performance figures are NVIDIA&amp;#39;s own.&lt;/p&gt;</content:encoded><category>Brains and Parts</category></item><item><title>Ant Group&apos;s Robbyant releases LingBot-Touch tactile model</title><link>https://roboticsdaily.robotshop.com/capsules/robbyant-lingbot-touch-tactile-model-2026-11/</link><guid isPermaLink="true">https://roboticsdaily.robotshop.com/capsules/robbyant-lingbot-touch-tactile-model-2026-11/</guid><description>LingBot-Touch is a tactile foundation model for robot hands, trained on nearly 5,000 hours of recordings from three sensor types, with weights and inference code public.</description><pubDate>Tue, 10 Nov 2026 00:00:00 GMT</pubDate><content:encoded>&lt;figure&gt;&lt;img src=&quot;https://roboticsdaily.robotshop.com/_astro/robbyant-lingbot-touch-tactile-model-2026-11.Cd-Q9gB3_1Bza3.jpeg&quot; alt=&quot;A silver-and-white dexterous robot hand with a wrist-mounted depth camera reaches toward a tabletop of varied objects, including a clear acrylic cube holding an SD card, a padlock, a small plant and a folded towel.&quot; /&gt;&lt;figcaption&gt;Image: TechNode / Robbyant · press-kit&lt;/figcaption&gt;&lt;/figure&gt;&lt;p&gt;Robbyant, the embodied-AI unit of Ant Group, released LingBot-Touch, a tactile foundation model for robot hands, with an arXiv paper dated 9 November 2026 and model weights and inference code made public, according to TechNode. The paper describes a unified representation across piezoresistive, magnetic and visuotactile sensors, intended to extend to new sensor types.&lt;/p&gt;&lt;p&gt;The model&amp;#39;s tactile tokens are presented as plug-and-play for different policy types. The authors test them with a diffusion policy and with the vision-language-action models pi0.5 and LingBot-VLA, in simulation and on real robots across several hardware platforms. TechNode reports pretraining on nearly 5,000 hours of recordings from three tactile sensor types, and evaluation on 26 real-world tasks with two robotic hands, including handling fragile objects and coordinating both hands.&lt;/p&gt;&lt;p&gt;Weights for lingbot-touch-tiny and lingbot-touch-small are listed on Hugging Face; the small model&amp;#39;s card states an Apache 2.0 licence. The paper is a preprint, and its performance claims are the authors&amp;#39; own and have not been independently reproduced.&lt;/p&gt;</content:encoded><category>Brains and Parts</category></item><item><title>XPeng and academic partners release GroundingPI, an open-source vision model for robots</title><link>https://roboticsdaily.robotshop.com/capsules/xpeng-groundingpi-open-source-perception-2026-11/</link><guid isPermaLink="true">https://roboticsdaily.robotshop.com/capsules/xpeng-groundingpi-open-source-perception-2026-11/</guid><description>On 4 November 2026 XPeng and academic partners released GroundingPI, a 4-billion-parameter open-source perception model for physical AI, with weights and code public.</description><pubDate>Sun, 08 Nov 2026 00:00:00 GMT</pubDate><content:encoded>&lt;figure&gt;&lt;img src=&quot;https://roboticsdaily.robotshop.com/_astro/xpeng-groundingpi-open-source-perception-2026-11.BjwYN7KB_ZJT5J7.jpeg&quot; alt=&quot;A collage of GroundingPI visual-perception demos on a purple-bordered board, showing labelled detections on road traffic, street signs, a night market, robot-arm scenes, zebras, penguins and a humanoid robot reading a store shelf.&quot; /&gt;&lt;figcaption&gt;Image: XPENG / press room · press-kit&lt;/figcaption&gt;&lt;/figure&gt;&lt;p&gt;XPeng and academic partners introduced GroundingPI on 4 November 2026, an open-source visual perception foundation model for physical AI. It locates objects and their positions as points and bounding boxes, and is trained across 11 perception capability categories. The model has about 4 billion parameters. Its paper, weights and code are public, and the release names Peking University and the University of Hong Kong among the partners.&lt;/p&gt;&lt;p&gt;In the team&amp;#39;s own evaluation, GroundingPI scored 73.68 on an equally weighted aggregate across 34 public perception benchmarks, against 71.54 for GPT-6 Astra, with 44 baseline models compared. In the RoboCasa-GR1 simulation, robot learning systems using it as their visual backbone reached higher task success with 50% of the demonstration data than comparison models did with 75%.&lt;/p&gt;&lt;p&gt;XPeng says the comparison covers perception benchmarks only and does not assess general intelligence or reasoning. It calls the research exploratory, notes that downstream results come from simulation and offline evaluation, and says the release is not an announcement of integration into XPeng products, including its IRON humanoid. The release does not name a specific open-source licence.&lt;/p&gt;</content:encoded><category>Brains and Parts</category></item><item><title>Infineon launches AURIX RC1, its first RISC-V microcontroller platform</title><link>https://roboticsdaily.robotshop.com/capsules/infineon-aurix-rc1-riscv-2026-11/</link><guid isPermaLink="true">https://roboticsdaily.robotshop.com/capsules/infineon-aurix-rc1-riscv-2026-11/</guid><description>The platform offers up to 12 cores and an edge-AI unit, with samples targeted for the first half of 2027.</description><pubDate>Thu, 05 Nov 2026 00:00:00 GMT</pubDate><content:encoded>&lt;figure&gt;&lt;img src=&quot;https://roboticsdaily.robotshop.com/_astro/infineon-aurix-rc1-riscv-2026-11.B5AzzCHv_2jqtMl.jpeg&quot; alt=&quot;A rendered Infineon AURIX RC1 microcontroller chip, marked with the Infineon logo, floating above translucent layers of binary code against a teal background with light streaks.&quot; /&gt;&lt;figcaption&gt;Image: Infineon Technologies / product page · press-kit&lt;/figcaption&gt;&lt;/figure&gt;&lt;p&gt;Infineon announced on 4 November 2026 AURIX RC1, its first microcontroller platform built on the open RISC-V architecture. It offers up to 12 32-bit cores, a parallel processing unit for edge AI and up to 40 MB of embedded memory, and is designed to support applications up to the ASIL D automotive safety level.&lt;/p&gt;&lt;p&gt;Infineon plans a later high-end lineup with 64-bit microprocessor capabilities, which would make the family a hybrid of microcontroller and microprocessor. The platform supports CAN-XL, 10BASE-T1S and multi-gigabit Ethernet, and the company says it meets ISO/SAE 21434 and the EU Cyber Resilience Act. Samples are targeted for the first half of 2027, and the chip will be shown at electronica in Munich from 10 to 13 November.&lt;/p&gt;&lt;p&gt;The stated target markets are vehicle zone, body and chassis control, battery management, powertrains, driver-assistance safety management and cockpits. Infineon says its ambition is to extend the platform&amp;#39;s real-time control to physical AI, but robotics is named only in the context of its trade-show presence, and no samples are shipping yet.&lt;/p&gt;</content:encoded><category>Brains and Parts</category></item><item><title>A preprint trains 96 real robots on pooled experience and lifts phone packaging to 98.3%</title><link>https://roboticsdaily.robotshop.com/capsules/tsinghua-agibot-cel-96-robots-2026/</link><guid isPermaLink="true">https://roboticsdaily.robotshop.com/capsules/tsinghua-agibot-cel-96-robots-2026/</guid><description>A 44-author arXiv preprint posted on 19 October 2026 reports a fleet of 96 robots sharing execution data to improve one policy.</description><pubDate>Fri, 30 Oct 2026 00:00:00 GMT</pubDate><content:encoded>&lt;figure&gt;&lt;img src=&quot;https://roboticsdaily.robotshop.com/_astro/tsinghua-agibot-cel-96-robots-2026.5XnhdsDb_ZMwR1S.jpeg&quot; alt=&quot;A wide hall with rows of identical white dual-arm humanoid robots seated at identical desks on a green floor under long strip lights, with a rose arranged in a vase at the front robot&amp;#39;s workstation.&quot; /&gt;&lt;figcaption&gt;Image: Yu et al., arXiv 2610.25736 (CC BY 4.0) · cc-by&lt;/figcaption&gt;&lt;/figure&gt;&lt;p&gt;A 44-author preprint posted to arXiv on 19 October 2026, with a second version on 21 October, describes Collective Experience Learning, a method in which deployed robots pool their own execution experience to update a shared policy. It aims to reduce reliance on expert corrections and needs no additional expert demonstrations.&lt;/p&gt;&lt;p&gt;The technique, called failure-guided experience retrieval fine-tuning, retrieves segments from successful trajectories based on the context of failures and uses them for supervised updates. Across five real-world manipulation tasks, the authors report success rates of 93.3% to 99.2%, measured over 120 trials per task on robots that were not used to collect data.&lt;/p&gt;&lt;p&gt;On multi-stage phone packaging, a fleet of 96 robots raised success from 79.2% to 98.3%. With collection time per robot held fixed, larger fleets generally performed better across the fleet sizes tested.&lt;/p&gt;&lt;p&gt;This is a preprint and has not been peer reviewed. The results are the authors&amp;#39; own, and the abstract page does not name the robot model or the affiliations.&lt;/p&gt;</content:encoded><category>Brains and Parts</category></item><item><title>SAP and Intel link warehouse software to autonomous robots</title><link>https://roboticsdaily.robotshop.com/capsules/sap-intel-physical-ai-warehouse-poc-2026/</link><guid isPermaLink="true">https://roboticsdaily.robotshop.com/capsules/sap-intel-physical-ai-warehouse-poc-2026/</guid><description>At SAP TechEd on 27 October 2026, SAP showed a proof of concept at Intel&apos;s Chengdu site that turns warehouse transport tasks into robot missions.</description><pubDate>Thu, 29 Oct 2026 00:00:00 GMT</pubDate><content:encoded>&lt;figure&gt;&lt;img src=&quot;https://roboticsdaily.robotshop.com/_astro/sap-intel-physical-ai-warehouse-poc-2026.CFAUoGFE_Z1AKRgw.jpeg&quot; alt=&quot;A title card for the SAP TechEd 2026 demo Physical AI by SAP: Robots in Action, with an SAP executive on stage next to a humanoid robot.&quot; /&gt;&lt;figcaption&gt;Image: SAP / YouTube · video-thumbnail&lt;/figcaption&gt;&lt;/figure&gt;&lt;p&gt;SAP announced its physical AI offering on 27 October 2026 at SAP TechEd, with a proof of concept built with Intel. Transport tasks created in SAP Extended Warehouse Management are translated into robot missions and carried out autonomously, using Intel&amp;#39;s edge AI and robotics foundation. Anu Intelligence is named as an ecosystem partner.&lt;/p&gt;&lt;p&gt;The proof of concept ran at Intel&amp;#39;s Chengdu site in China. SAP says material movement was automated end to end, reducing both transport distances and repetitive manual handling, but the announcement gives no figures.&lt;/p&gt;&lt;p&gt;SAP describes the approach as vendor and hardware agnostic: customers choose their own robots, devices and AI models, while SAP provides business context, orchestration and governance. That is the company&amp;#39;s claim; the announcement cites no independent test of compatibility across robot makers.&lt;/p&gt;</content:encoded><category>Brains and Parts</category></item><item><title>Percepto launches Axion, a hardware-agnostic platform for inspection data</title><link>https://roboticsdaily.robotshop.com/capsules/percepto-axion-platform-launch-2026/</link><guid isPermaLink="true">https://roboticsdaily.robotshop.com/capsules/percepto-axion-platform-launch-2026/</guid><description>The inspection-drone company&apos;s new software service runs more than 160 AI models on data from third-party drones, sensors and archives, announced on 27 October 2026.</description><pubDate>Wed, 28 Oct 2026 00:00:00 GMT</pubDate><content:encoded>&lt;figure&gt;&lt;img src=&quot;https://roboticsdaily.robotshop.com/_astro/percepto-axion-platform-launch-2026.COEVnYfR_Z1taBuE.jpeg&quot; alt=&quot;A dark Axion platform screen reading Harnessing AI to turn infrastructure inspection data into Insights, with a drag-and-drop upload area and a Connect your data label.&quot; /&gt;&lt;figcaption&gt;Image: Percepto Physical AI for Critical Operations / YouTube · video-thumbnail&lt;/figcaption&gt;&lt;/figure&gt;&lt;p&gt;Percepto, an Austin, Texas inspection-drone company, launched Axion on 27 October 2026, according to DroneLife. Axion is a software-as-a-service platform for infrastructure operators that accepts inspection data from third-party drones, sensors and archives, not only from Percepto&amp;#39;s own hardware.&lt;/p&gt;&lt;p&gt;The company says the platform runs more than 160 AI models, trained on more than 100,000 hours of real-world autonomous inspection data gathered over more than a decade. It calls Axion a physical AI platform: the models turn raw inspection captures into findings that operators can track over time.&lt;/p&gt;&lt;p&gt;The launch comes with a new leadership team: Julien Bellanger as CEO, Kurt Nantkes as chief revenue officer and Amir Khoshniyati as chief marketing officer. The article names no Axion customers; it mentions only an unnamed U.S. energy company that gave Percepto a multi-year transmission line inspection contract. The model and data figures are the company&amp;#39;s own.&lt;/p&gt;</content:encoded><category>Brains and Parts</category></item><item><title>NVIDIA adds open tools to train, test and deploy its Cosmos world models for robots</title><link>https://roboticsdaily.robotshop.com/capsules/nvidia-cosmos-open-tools-world-models-2026-10/</link><guid isPermaLink="true">https://roboticsdaily.robotshop.com/capsules/nvidia-cosmos-open-tools-world-models-2026-10/</guid><description>NVIDIA released Cosmos Data Factory tools and blueprints on 21 October 2026, plus smaller Cosmos 3 Nano models on Hugging Face and faster inference services.</description><pubDate>Tue, 27 Oct 2026 00:00:00 GMT</pubDate><content:encoded>&lt;figure&gt;&lt;img src=&quot;https://roboticsdaily.robotshop.com/_astro/nvidia-cosmos-open-tools-world-models-2026-10.D4rLBVkR_29tx3v.jpeg&quot; alt=&quot;A three-panel composite of world-model-generated scenes from NVIDIA Cosmos: a worker crossing a factory yard as a forklift carries pallets, a white humanoid robot reaching out in a lab, and a grey sports car pulling out of a parking lot.&quot; /&gt;&lt;figcaption&gt;Image: NVIDIA / NVIDIA Blog · press-kit&lt;/figcaption&gt;&lt;/figure&gt;&lt;p&gt;NVIDIA announced on 21 October 2026 a set of new tools for Cosmos, its family of world models used to generate and evaluate training data for robots and autonomous vehicles. Cosmos Data Factory provides tools for labelling, curating, augmenting and evaluating training data, with five ready-made workflows, and a Cosmos Framework packages these steps as skills that agents can use with other open models, including GR00T and Pi0.5.&lt;/p&gt;&lt;p&gt;NVIDIA also published Cosmos 3 Nano simulation variants on Hugging Face for testing autonomous policies, and updated its Cosmos 3 inference services, which it says run up to twice as fast with no loss of quality in validated configurations. Named adopters include Porsche, which NVIDIA says needs up to 90% fewer repeated physical test drives, and robot maker ANYbotics.&lt;/p&gt;&lt;p&gt;The licensing terms were not stated on the announcement, and the adopter figures are NVIDIA&amp;#39;s own. Generated and curated data is becoming as central to robot development as the robots themselves.&lt;/p&gt;</content:encoded><category>Brains and Parts</category></item><item><title>Omron agrees to buy industrial-computer maker OnLogic for $675 million</title><link>https://roboticsdaily.robotshop.com/capsules/omron-onlogic-acquisition-2026/</link><guid isPermaLink="true">https://roboticsdaily.robotshop.com/capsules/omron-onlogic-acquisition-2026/</guid><description>Omron signed a definitive agreement in October 2026 to acquire OnLogic, a U.S. maker of industrial edge computers, for $675 million, with closing expected around January 2027.</description><pubDate>Tue, 27 Oct 2026 00:00:00 GMT</pubDate><content:encoded>&lt;figure&gt;&lt;img src=&quot;https://roboticsdaily.robotshop.com/_astro/omron-onlogic-acquisition-2026.C0WsWXbp_1VQvK6.jpeg&quot; alt=&quot;A group of five orange-and-grey OnLogic fanless industrial computers of different sizes, with their front-panel USB, Ethernet and display ports visible, arranged on a plain orange background.&quot; /&gt;&lt;figcaption&gt;Image: OnLogic / product page · press-kit&lt;/figcaption&gt;&lt;/figure&gt;&lt;p&gt;Japanese automation company Omron signed a definitive agreement to acquire all of OnLogic, a maker of industrial and rugged edge computers based in South Burlington, Vermont, OnLogic announced on 13 October 2026. Reuters, as published by The Star, put the price at $675 million, covering OnLogic and its parent company, Interlogic.&lt;/p&gt;&lt;p&gt;Closing is expected in or around January 2027, subject to regulatory approval. OnLogic, founded in 2003 with more than 70,000 customers, will keep operating as an independent business with its brand and its chief executive, Sean Larkin. Omron plans to combine OnLogic&amp;#39;s computers with its own control technology to build platforms for physical AI on the factory floor.&lt;/p&gt;&lt;p&gt;OnLogic&amp;#39;s revenue and headcount were not disclosed. Running AI models next to machines requires rugged computers in the cell itself, and the deal shows automation suppliers moving to own that layer.&lt;/p&gt;</content:encoded><category>Brains and Parts</category></item><item><title>AMD agrees to acquire Fei-Fei Li&apos;s World Labs, a maker of world models, for $8.2 billion</title><link>https://roboticsdaily.robotshop.com/capsules/amd-world-labs-acquisition-2026/</link><guid isPermaLink="true">https://roboticsdaily.robotshop.com/capsules/amd-world-labs-acquisition-2026/</guid><description>AMD said on 28 September 2026 it will acquire World Labs, which builds AI models of physical environments, for $8.2 billion; founder Fei-Fei Li joins AMD as chief scientist.</description><pubDate>Tue, 27 Oct 2026 00:00:00 GMT</pubDate><content:encoded>&lt;figure&gt;&lt;img src=&quot;https://roboticsdaily.robotshop.com/_astro/amd-world-labs-acquisition-2026.BEiAUmNB_1S5TOb.jpeg&quot; alt=&quot;Lisa Su and Fei-Fei Li seated side by side in a Bloomberg studio, Su gesturing as she speaks.&quot; /&gt;&lt;figcaption&gt;Image: Bloomberg Television / YouTube · video-thumbnail&lt;/figcaption&gt;&lt;/figure&gt;&lt;p&gt;AMD announced on 28 September 2026 that it will acquire World Labs, the company founded by Stanford computer scientist Fei-Fei Li, for $8.2 billion, TechCrunch reported. The deal is expected to close before the end of 2026, subject to regulatory approval. Li will join AMD as executive vice president and chief scientist.&lt;/p&gt;&lt;p&gt;World Labs, founded in 2024, builds world models: AI systems that generate and reason about three-dimensional physical environments. Its first product, Marble, creates explorable 3D worlds for entertainment and for simulated environments used to train robots. AMD and World Labs already had a partnership on optimizing the training and inference of these models.&lt;/p&gt;&lt;p&gt;The form of payment and AMD&amp;#39;s product plans for the technology were not disclosed. Simulated worlds are one of the main ways robot models are trained before they touch real hardware, and chipmakers are buying their way into that layer of the stack.&lt;/p&gt;</content:encoded><category>Brains and Parts</category></item></channel></rss>