"Meet ViserDex: a robust visual sim-to-real approach for dexterous in-hand reorientation! 🖐️🤖We propose a new RGB-based visual sim-to-real pipeline that bridges the gap between perception and control. By performing domain randomization directly on 3D Gaussians, we generate photorealistic training data that transfers seamlessly to the real world. The result? A physical multi-fingered hand achieving an average of 25+ consecutive reorientations—even under extreme visual conditions that break traditional pose estimators!" 🔗 Project Website: https://lnkd.in/eUef4sGZ... 📄 paper ETH Zurich, NVIDIA, in comment 📍 Authors: Arjun Bhardwaj, Maximum Wilder-Smith, Mayank Mittal, Vaishakh Patil, Marco Hutter ViserDex: Visual Sim-to-Real for Robust Dexterous In-hand Reorientation Robotic Systems Lab, ETH Zürich https://lnkd.in/e8ErGKiE

The scaled and actor controlled scenario generation & variation enabling training the brains ML (models) of AV technology.

Reverse engineering is evolving beyond manual reconstruction. We are developing an AI-driven tool focused on transforming scan data and point clouds into intelligent 3D reconstruction workflows. The ability to extract usable structure from complex spatial data opens new possibilities for reconstruction, analysis, and digital asset generation at scale. Point clouds are no longer just raw data — they’re becoming interpretable geometry. Coming soon to the FutureSnap platform #ReverseEngineering #PointCloud #3DScanning #3DReconstruction #SpatialComputing

200 milliseconds. That is the response cycle of the new interaction model from Thinking Machines Lab. The model processes audio, video, and text simultaneously, every 200ms, in a continuous loop. Benchmarks: Interactivity: 77.8 on (GPT Realtime: 46.8) Time awareness: 64.7%. (GPT Realtime: 4.3%) Interactivity isn't a harness; it’s baked into the weights. It’s the first time a model has demonstrated the "reflexes" to interrupt a human based on a visual event. No existing frontier model can do any of this. It is a realtime actual copilot.

Vision models fail when grid layouts shift to polar coordinates, with accuracy dropping from 83% to 39%. This reveals a reliance on textual shortcuts. arxiv.org/abs/2605.09883

I first came across sparse autoencoders from an AI-generated roundtable discussion. SA is powerful for understanding the black box inside the LLM and crucial for validating and optimizing LLM behavior. It is quite amazing that Qwen provides open access to their Qwen-Scope, showing the power of open-weight models. https://lnkd.in/epUifyi9

🚀 And AIPC.computer is the ONLY AI-powered marketplace which exists for these next-gen PCs.computer 💻 💻 Laptops.computer 🖥️ Desktops.computer 🖥️ Workstations.computer 🤝 Open to working a win-win deal: Michael Hahnel · Kevin Scott · Nicole Dezen · Judson Althoff 🌐 AI.commerce.computer — Discover our 12 AI-powered marketplaces built for AiAgents.computer ⚡ 🧠 Powered by: 🔲 Chip.computer 💾 Software.computer 🖥️ Hardware.computer #AIPC #AIcomputers #NextGenTech #ComputerMarketplace #Partnership #WinWin #AIEcosystem #AIcommerce #AIAgents #AIMarketplaces #ChipComputer #SoftwareComputer #HardwareComputer

Six months ago, "TOPS" was the only number on every AI-PC pitch. Six months later, battery is the only number that decides if a feature ships. Compute is the marketing budget. Battery is the engineering budget. The carousel walks through the math nobody shows on a spec sheet: → why Surface users judge AI by hours, not TOPS → how Surface Laptop and Surface Pro scenarios change the battery story → the three knobs operators turn (most teams only turn one) Examples are public product categories, not internal Surface measurements. What's your team's battery-per-AI-feature target? #AIPC #CopilotPC #Battery #ProductManagement #Surface

One uncomfortable truth about computer vision: Your model starts getting outdated the moment you deploy it. Not because the model is bad — but because the world keeps changing: * new environments * new edge cases * new object variations * new camera setups But most vision systems are still built like this: Train once → Deploy → Hope it works Reality is different: **Data is continuously evolving. Models are not.** Maybe the real problem isn’t model accuracy — it’s the lack of continuous adaptation. Curious: How do you handle “model drift” in real-world vision systems? Or do most systems just silently degrade over time?

We’re experimenting with automated 3D semantic annotation for 3DplusA digital twins using VLMs. The workflow is simple: - Click on any asset, - Search for any asset by text, or - Select predefined categories, and the system instantly identifies it in the 3D environment.