Embedded Systems

📺 Creating the invisible earphones from scratch The author built a pair of custom invisible wireless earphones by completely modifying a pair of wireless earbuds. One earphone was molded to fit inside the ear using silicone and clear resin, while the second was miniaturized and built into a pen as a tiny wireless microphone. The original electronics, touch controls, and charging case were modified to fit the new designs. The project combines reverse engineering, electronics miniaturization, soldering, custom ear molding, resin casting, and precision assembly. ⛓ https://youtu.be/q7nG6rCxbK4 #video #projects

📱📱 Samsung Firmware Update Bricks Smart Refrigerators Samsung’s Bespoke AI smart refrigerators experienced failures after the latest software update. Affected units, reportedly including many four-door models from 2024 or later, suddenly lost power or stopped functioning after the update. Some displays became stuck on an update-related message, while SmartThings showed the refrigerators as offline. Samsung said the incident was caused by an error during internal testing of a refrigerator software update and that it suspended the testing after receiving reports. Samsung said the issue is limited to South Korea and has activated an emergency response through its service centers. Korean media reported hundreds of customer complaints, with some owners saying their refrigerators stopped cooling and food had to be discarded. Samsung has not disclosed how many units were affected or what measures will prevent a recurrence. ⛓ More information... #news

🤖 Influencer Takes on Six-Foot Humanoid Robot The 📺 video billed as the first human-versus-robot MMA fight shows influencer Frankie LaPenna facing a six-foot T800 humanoid robot developed by Chinese company. The event took place in 🇺🇸 San Francisco on 18th of September. The dramatic footage quickly spread online, though questions were raised about how authentic the fight was. Developed by China-based EngineAI, the robot stands about 1.73 meters tall, weighs up to 85 kg and has 29 degrees of freedom. The released video shows LaPenna lands several punches before the robot responds with powerful kicks that eventually knock him to the floor. REK later confirmed that the robot was controlled remotely by a human rather than fighting autonomously, and said LaPenna continued the match and ultimately defeated the robot. More footage is expected to be released. ⛓ More information... #news #robots

🧠 Robots Get a Sense of Touch Robots are getting better at complex tasks with vision-language-action (VLA) models, but they still struggle with precise manipulation such as inserting cables, turning keys, or handling deformable objects. Tactile sensing can provide information about force, grip, and slipping, but progress has been limited by the lack of large, diverse datasets. Recent research is addressing this gap by combining touch with vision and robot-action data. Berkeley researchers created the T-Rex dataset with about 100 hours of tactile demonstrations involving more than 200 household objects. Another project aggregated over 3,000 hours of publicly available data covering 21 sensor types and different robot platforms. USC researchers also released a model that can infer tactile information from visual data, potentially helping researchers create more tactile training data from existing vision datasets. Early results show that touch can significantly improve robot manipulation. In one study, tactile feedback produced a 65% average success rate across 12 tasks, nearly twice the performance of the best VLA baseline; another approach achieved 62.8% success compared with 28.2% without touch. The main challenges remain data diversity, sensor compatibility, and how robots process sparse tactile signals in real time. ⛓ More information... #articles #robots #artificial_intelligence

🧠 AI Image Generation on an RP2350 Microcontroller This project brings generative AI image generation to a $1 RP2350 microcontroller, as used in the 🍓 Raspberry Pi Pico 2. Its compact model and inference code fit within 4 MB and run using 520 KB of RAM, generating 128×128 RGB face images in roughly 10–20 seconds. Features of the project: ➡️Latent flow diffusion transformer ➡️Five-class conditional generation ➡️Int8 post-training quantization ➡️VGA and USB output The project demonstrates how far generative AI can be pushed on extremely constrained hardware, combining model compression with a highly optimized inference engine. The source code is available on the dedicated 😹 GitHub repository alongside with examples. ⛓ More information... #project #raspberry #artificial_intelligence