General Intuition, a Bezos-backed startup valued at $2.3 billion, argues that video game data is superior to internet data for training AI that understands physical space and time. CEO Pim de Witte explains on TechCrunch's Equity podcast how world models trained on gaming data could advance physical AI and robotics, how the company evolved from gaming platform Medal TV, and the ethical considerations around potential defense applications. The company recently closed a $320 million round with investors including Coatue, Eric Schmidt, and researchers from MIT and Google DeepMind.
Nguồn: https://techcrunch.com/video/why-this-ceo-thinks-video-games-make-better-training-data-than-the-internet. 8sync News chỉ tóm tắt và dẫn link; bản quyền nội dung thuộc tác giả và nguồn gốc.
Dự án CPSDrone tạo ra những chú cá robot cỡ nhỏ cho bể thủy sinh, giải quyết các thách thức như truyền thông vô tuyến dưới nước ở tần số 433 MHz, vỏ chống nước in 3D bằng công nghệ seal vòng O-ring, và sạc tự động qua trạm docking. Những robot này được điều khiển tập trung thay vì hoạt động hoàn toàn tự chủ, trong khi thuật toán boids có thể thay thế vai trò điều khiển trung tâm nhờ tính đơn giản và khả năng mô phỏng đàn cá sinh động.
Đọc bài này để khám phá cách giải quyết những thách thức kỹ thuật thực tế trong thiết kế các robot dưới nước nhỏ gọn, từ giao tiếp vô tuyến dưới nước đến tối ưu hóa hệ thống tự động hóa thông qua các giải pháp đơn giản như thuật toán boids.
China's 15th five-year plan (2026-2030) marks a strategic pivot from chip manufacturing to AI deployment and export. The plan mentions AI over 50 times, designates robotics as a strategic industry, and sets out a global agenda including a World AI Cooperation Organisation and Belt and Road AI platforms targeting Global South nations. For South Africa, this creates both opportunity — cheaper clean energy, BRICS computing access, smart-manufacturing partnerships — and risk. South African solar imports from China jumped 81.4% year-on-year, yet China scrapped its solar export tax rebate, threatening price increases. The foundational AI and hardware layers remain in China while South Africa would receive only the application layer. Analysts warn that countries faring best in US-China AI competition are those that articulate their own AI priorities early and negotiate from a position of strong data protection rules — a challenge South Africa's still-draft AI policy has yet to meet.
Forterra has deployed over 100 autonomous Lancer ATVs in Ukraine over the past nine months, marking what the company claims is the largest deployment of US autonomous ground vehicles in combat. The gas-powered vehicles, based on Polaris ATVs with custom sensors and compute, have completed 1,100+ missions, driven 2,500+ miles, and carried over 777,000 pounds of cargo including 52 casualty evacuations. Ukrainian soldiers have largely been teleoperating rather than fully autonomously running the vehicles, as current autonomy cannot yet react to live enemy threats. Key lessons learned include electronic warfare adaptation, remote software updates, and terrain navigation. Forterra is now combining classical robotics with generative AI to close the gap toward full battlefield autonomy, while facing pressure to reduce costs as attrition on the battlefield is significant.
General Intuition, a startup backed by Vinod Khosla, has raised $320 million at a $2.3 billion valuation on the thesis that video game data can train foundation models for physical AI. CEO Pim de Witte argues that just as GPT-style models replaced task-specific NLP models, a general-purpose foundation model for embodied AI will eliminate the need for massive real-world robotics datasets. The company demonstrated its model powering a quadrupedal robot after only eight minutes of real-world fine-tuning. General Intuition's goal is not to build robots itself, but to become the foundational layer that other robotics companies build upon.
MIT and Tecnológico de Monterrey have developed FrED (Fiber Extrusion Device), a low-cost desktop manufacturing device built and operated by students in an educational factory setting. The initiative replaces textbook learning with hands-on factory experience, training nearly 500 students in advanced manufacturing automation across campuses in Cambridge, Monterrey, and Mexico City. FrED generates multi-modal data suited for digital twins, AI-driven process improvement, and smart manufacturing education aligned with Industry 4.0. A new FrED factory is opening at Tec's Saltillo campus, with plans to expand globally. The collaboration has produced 25 publications, an ASEE Best Paper Award, and a new undergraduate research curriculum called FRAME.
NVIDIA and Hugging Face are integrating NVIDIA Isaac GR00T 1.7 (an open VLA model for humanoid robots) and Isaac Teleop (a data collection framework) into LeRobot, Hugging Face's open source robotics library. NVIDIA Cosmos 3, a frontier world foundation model, is also planned for future integration. Together these additions give developers a standardized, open pipeline for robot data collection, model training, fine-tuning, simulation, and deployment. The collaboration connects NVIDIA's 3 million robotics developers with Hugging Face's 16 million AI builders, and builds on existing resources including a large open physical AI dataset with over 15 million downloads, Isaac Sim/Lab simulation frameworks, and Jetson Thor hardware integration.
Nomagic, a Warsaw-based warehouse robotics company, has deployed a vision-language-action (VLA) model in live customer warehouses, claiming it roughly halved robot-caused human interventions. Unlike most robotics labs racing to build general-purpose robot brains, Nomagic bets on task mastery first, wrapping its imperfect VLA in classical safety software to meet the 99.9% reliability threshold required in real operations. The company trains its models on live production data — including millions of picks monthly from Zalando — rather than simulations. Its first VLA deployment runs at Brack.Alltron, Switzerland's second-largest e-commerce platform, enabling autonomous overnight and weekend shifts. The AI lab is led by Markus Wulfmeier, a former Google DeepMind and Gemini Robotics researcher.
A social media round-up from RoboCup 2026, held in Incheon, South Korea from July 2-6. The event featured robotics competitions, training sessions, and a symposium, with highlights gathered from participants' social media posts.