Javier Morales @MLRobotTech

Dense to MoE Adaptation for Compact Vision Language Action Policies Muchun Niu, Shuang Chen, Yuzhou Wu, Linfeng Zhang arxiv.org/abs/2609.16503 [𝚌𝚜.𝚁𝙾]

El coste de la máquina parada: dónde recortar mantenimiento sin pagarlo después ift.tt/GTu910W

100%. Even in the AI era, data, customer lock-in and distribution are still the same entry barriers.

Codez @0xCodez

2 days ago

i don't think people realize what this unlocks... Tesla's real moat was never the cars. Karpathy spent years saying it out loud: the model doesn't get better from clever architecture, it gets better from millions of vehicles streaming video and telemetry back home.

15 65 560 101K 418

AIs cracking short-term superforecasting is such an important (and, still, under-discussed) trend. Via a new piece from @nikostro economist.com/science-and-te…

@Mihaildasp @360AirServices thanks for following us! Visit mundocompresor.com to check out the latest news. We invite you to subscribe to receive all the information and news mundocompresor.com/suscripcion-gr… Great to connect !

ipXchange's Edge AI in Production Week is less than two weeks away! Join Arduino, @NXP, @ExpansoIO, and @Advantech_USA for a week of engineer-led sessions on what it actually takes to build, deploy, and scale production-ready edge AI systems. Register: luma.com/kp5viqoy?utm_s…

This is where 3D generation gets really interesting. Going from a single image to an actual 3D scene with an agent handling the workflow feels like a much bigger shift than just “better 3D generation.” The agent + MCP combination is the part I’m most curious about.

Hyper3D by Deemos @DeemosTech

a week ago

One image. One Agent. One scene. @OpenAIDevs GPT-6 Astra (Codex) + HYPER3D MCP just showed us the second half of 3D gen. 🚀

40 217 2K 349K 2K

CM-YOLO: A Multimodal PCB Defect Detection Method Based on Cross-Modal Feature Fusion mdpi.com/1424-8220/25/1… #PCB #defect_detection #multimodal #feature_fusion

Ansaldo Green Tech has opened a high-volume AEM electrolyser stack production line at its Genoa facility in Italy, with annual capacity of 300MW. The line will produce 1MW stacks for modular green hydrogen systems. fcw.sh/9DOfgJ #Hydrogen #Electrolysers

Universal Robots y MiR muestran cómo automatizar la intralogística sin perder flexibilidad mundocompresor.com/noticias-de-ev… #automatización #intralogística

GPT-6 Astra can become something bigger than an AI agent: an AI Operating System. the formula: AI OS = Reasoning + Execution + Memory + Router + Tools + Verification not an AI that simply answers prompts. a system that can understand a goal, break it down, execute the work, inspect the result and continue autonomously. step 1 -> GPT-6 Astra becomes the reasoning layer. complex problems, planning, architecture and decision-making happen here. its job is to understand the objective and turn ambiguity into an actionable plan. step 2 -> give Astra an execution layer. it writes code, operates tools, creates artifacts, controls applications and turns decisions into finished work. step 3 -> build a persistent state layer. store objectives, decisions, constraints, previous attempts, errors and next actions outside the chat. Astra should always know what happened and where the workflow should continue. step 4 -> add the router. different tasks require different tools, environments and execution paths. the router decides what should happen next and sends each task to the right capability. step 5 -> create the execution loop: plan -> execute -> inspect -> update state -> continue if something fails, the system doesn't restart from zero. it learns from the failure, updates the state and continues from the exact point where it stopped. step 6 -> verify before completion. tests, source checks, constraints and explicit success criteria decide whether the work is actually finished. if not, the system loops back and fixes the failure. that is the difference between using an AI model and building an AI Operating System. GPT-6 Astra doesn't just generate answers. it can reason, build, operate tools, inspect results and keep moving toward the objective. one prompt gives you an answer. a persistent system gives you a workflow that can actually run.

hammertime @hammertime_one

2 weeks ago

x.com/i/article/2096…

9 2 50 28K 23

Google dropped this 145 pages documenting how researchers use Gemini to tackle scientific problems. 𝘚𝘢𝘷𝘦 & 𝘙𝘦𝘵𝘸𝘦𝘦𝘵 (𝘵𝘰 𝘩𝘦𝘭𝘱 𝘺𝘰𝘶𝘳 𝘯𝘦𝘵𝘸𝘰𝘳𝘬) A few things that stood out to me (in simple terms): - In one case, the AI was used as an adversarial reviewer and caught a serious flaw in a cryptography proof that had passed human review. That’s a very different use than “summarise this PDF.” - The model links tools from very different fields (for example, using theorems from geometry/measure theory to make progress on algorithms questions). This is where its wide reading really matters. - They don’t let the model run wild. Humans still choose the problems, check every proof, and decide what’s actually new. The model is there to suggest ideas, spot gaps, and do the heavy algebra. - Agentic loops, not just chat In some projects, they plug Gemini into a loop where it: -- proposes a mathematical expression, -- writes code to test it, -- reads the error messages, and -- fixes itself. (humans only step in when something promising appears) We are moving past the era of simple chat prompts and into a more sophisticated era of research. ⮑ If your institution is interested in hosting an AI session or a workshop, request your training here: forms.gle/dbRtc7j2W4zZyL…

11 years of humanoid robotics in one video. 🤖 In 2015, humanoid robots were still struggling with basic tasks like walking, climbing stairs and recovering from falls during the DARPA Robotics Challenge. Fast-forward to 2026, and humanoids are running, jumping, boxing, playing football, handling industrial tasks and operating in real-world environments. The World Humanoid Robot Games featured 2,056 robots across 51 events, including practical tasks in factories, hotels and homes. The progress is obvious. But the real question is not how impressive they look in a demo. What will humanoid robots be capable of in 2036? 👀 🎥 Media: Defense Advanced Research Projects Agency (DARPA) / Unitree Robotics ⚠️ This content is shared for informational purposes only. CTO Robotics Media is a media platform and does not own or develop the technologies shown. Credit belongs to the original creators. #HumanoidRobots #Robotics #PhysicalAI #AI #DARPA #RobotEvolution #FutureOfRobotics #EmbodiedAI #Automation #RobotTechnology #CTORoboticsMedia

Bio-inspired Learning and Decision-Making with Probabilistic In-Memory Computing Hardware: Part 2 Thomas Dalgaty, Eiji Kawasaki, Miguel de Prado, Tommaso Salvatori, Germain Haugou, Eric Flamand arxiv.org/abs/2609.11288 [𝚌𝚜.𝙰𝚁]

La respuesta de la UE ante la crisis arancelaria no convence al sector logístico y del packaging mundocompresor.com/especial-event… #empackmadrid #packaging @EmpackSpain

GPT-6 x Blender -> turned a simple 3D model into dark armor, a deep red cape, and a sunset battlefield. Now you can try it directly with PixVerse Plugin Full prompt ↓

100%. Simulation is rapidly becoming the default for general robot learning.

Siyuan Huang @siyuanhuang95

a week ago

Two months ago, at an RSS workshop panel, I said that general real2sim reconstruction is still far from solved, especially for articulated objects, which will be a major bottleneck for robotics. But now that has totally changed with Astra; using several photos as input, we can

keyboard_arrow_left Previous keyboard_arrow_right Next

7 21 163 19K 89

Data leakage can make a bad model look scientifically convincing. It happens when information from validation or test data quietly enters model development: scaling before the split, selecting features on the full dataset, tuning hyperparameters against the test set, or mixing correlated samples across train and test. The dangerous part is that nothing necessarily looks wrong. Your metrics improve. Your plots look cleaner. Your conclusions become more confident. But the model has partly seen the answer. A useful rule for scientific ML: Split first, then preprocess. Fit scalers, imputers, feature selectors, and dimensionality reduction on the training data only. Keep the test set completely untouched until the final evaluation. Data leakage is not just a coding mistake. It is a failure of experimental design. This is one of the practical issues I discuss in Inteligencia artificial para estudiantes de ciencias (Ediciones Pirámide).

Pensabas que la pintura a mano estaba a salvo?

thijs @cdngdev

2 weeks ago

i gave astra a robot, a paint brush, and a camera then asked it to paint the golden gate bridge in real life! it figured out how to control the robot, and progressively got better throughout its attempts. the timelapse is sick

667 2K 21K 4.9M 6K

AXIS Is Teaching Robots From Their Mistakes What if robots could learn not only how to complete tasks, but also how to recover when something goes wrong? @axisrobotics is exploring that idea through human corrections. A robot attempts a task independently. When it begins to fail, a human takes control, fixes the issue, and then hands control back. But not every correction is useful training data. A successful intervention may include pauses, unnecessary movements, or actions the robot did not actually need. AXIS therefore verifies each correction before using it for training. The results show how selective the process is: 660 corrections collected → 161 verified corrections used for training These verified short corrections improved performance across three evaluation seeds, increasing successful rollouts from 66 to 78/80/84 out of 160. The goal is not to make robots imitate every human movement. It is to identify the small action that actually fixes the failure. That creates a powerful learning loop: Robot fails → Human corrects → Useful correction is verified → Model learns → Robot improves AXIS V2 is focused not just on collecting more data, but on identifying which data genuinely helps robots improve.

Vinay @Spidyman00

2 weeks ago

Axis Challenger Program: Higher Difficulty, Higher Rewards @axisrobotics is introducing a new high difficulty task pool with 2–3× more points than standard tasks. But there’s a catch: access is whitelist only. Here’s how it works: → New Challenger tasks drop every Monday at

23 3 35 617 0