Eric Lee I BTS & SZF

Most data collection devices still suck in the same ways. · Camera module is too heavy on your head · Power bank gets hot · Sensors are hard to source · Annoying to put on and take off · Can’t swap the camera module · Data still isn’t clean enough We’ve been dealing with this with a lot of teams for a while, so we made EgoOne Lite to clean those up.

Chinese and Western data-glove development in 2026 is moving in different directions. Chinese systems this year have concentrated on denser tactile arrays, integrated motion-and-contact capture, and modular designs that make operator changes and maintenance easier. MX Matrix offers configurable fabric and EMF combinations; TachinGlove Snap separates electronics from fabric for faster swaps; THD1000R emphasizes three-axis force output. Western systems have leaned more toward precision motion, software integration, and feedback. Sharpa AE01 uses 22 encoders plus fingertip vibration; Melt ContactGlove3 Pro focuses on five-fingertip electromagnetic tracking with ROS 2 and Python SDKs; MANUS has prioritized SDK improvements and modular platform previews. The practical difference is less about which side has “better” sensors and more about what each prioritizes: Chinese products are pushing harder on contact density and operational practicality for data collection, while Western systems remain stronger on kinematic precision, established software interfaces, and teleoperation feedback. Neither fully replaces the other for most embodied AI pipelines. Full report: app.notion.com/p/shenzhenfou… #TactileSensing #RobotLearning #DexterousManipulation

We spent a few months researching the physical AI data collection market. There is a clear gap. Many products die at the prototype stage, and many more are likely to disappear next year. This pattern is not limited to data collection devices — it also appears in robots, wearables, and AI toys. The EgoOne series is what we built to address this gap in the data collection market. It’s structured to work with leading physical AI platforms, including @huggingface @LeRobotHF, NVIDIA Isaac/Cosmos, @Roboflow, @scale_AI, and @encord_team.

We put an RGB camera on a handheld UMI gripper and recorded a normal pick-and-place. After fixing the mount and improving marker detection, we got usable jaw-opening numbers on 98.7% of the frames in a 30-second clip (was ~95.3% before). Camera position relative to a table marker was available on ~84.8%. Not perfect, but good enough to actually look at what the hand was doing instead of just watching the video. 🧵1/6

Same intent. Zero collisions. ElRobot Follower, 7+1 DOF: “Put both pens in the box.” Approach, grasp, detour the obstacle, place, home — twice. The expensive loop is still try → break → fix. This workspace runs that loop in software first. Ghost preview, live J1–J8, plan that can leave the sim and hit the real arm. #PhysicalAI #ElRobot #ShenzhenFoundry

Packed. Labeled. Ready to fly. Weekend in the warehouse. Every order caught up. This is what “on time” actually looks like. Execution + Shenzhen speed. #shenzhen #APEC #physicalAI

Physical AI is drowning in demos. What’s still missing is unglamorous demand data: what teams will actually buy in the next 12 months, at what budget, and what’s still broken in current suppliers. If you build, buy, or evaluate robots/embodied AI hardware, this is a 2-minute industry survey. Name optional. Estimates are non-binding. What’s the one product category you’d actually write a PO for in 2026? Survey: docs.google.com/forms/d/1IsS… If you reply with your category — arms, grippers, data-capture, sensors, cameras, compute — I’ll share a summary of the results.

Walking’s solved. The real bottleneck for humanoids right now is the hands. We just dropped our latest report on dexterous hands as the data layer behind physical AI. Key takeaway: The strongest hands aren’t the ones with the highest DOF count. They’re the ones that turn contact into reliable data, explain their own failures, and improve without needing a full rebuild every time. We looked at 70+ hands, 30+ suppliers across Shenzhen and Dongguan, plus real buyer requirements. Bottom line? Test the full loop — mechanics, touch, timing, and data export — not just the spec sheet. What’s been your biggest headache with robot hands so far? Mechanics, data, or something else? Full report👇 app.notion.com/p/shenzhenfou… #HumanoidRobots #Robotics #DexterousHands

🇺🇸1X NEO is coming from full robot integration and real task capability. 🇨🇳Wuji is betting on back-drivable, biomimetic joints with strong independent control. 🇨🇳LinkerBot is optimizing for speed, precision, and complex motions that can scale. 🇨🇳Xynova is going lighter with hybrid drive and cleaner movement. Three of them are Chinese teams pushing these different approaches in parallel right now. The mechanics are moving fast. The real constraint isn’t finger motion anymore, it’s turning that motion into something you can actually control, get useful touch data from, and manufacture at stable production volumes without everything falling apart. #DexterousHands #HumanoidRobotics #PhysicalAI #EmbodiedAI #RoboticHands

VLA serves as a foundational paradigm in embodied AI, forming the basis for multimodal action models, where the accuracy and alignment of high-quality training data directly determine generalization and real-world deployment performance. This new headband data collector nails the basics: dual 1080P global shutter cams, wide FOV, 1kHz IMU, audio sync, lightweight and field-ready with open interfaces and customization. We’re assisting overseas clients navigating China’s market for all kinds of VLA hardware right now. Requirements are highly custom and non-standard, feels a lot like the earliest mobile power banks in the iPhone 4 days. But just as that era exploded the smartphone accessory world, AI is opening intelligentization across the entire electronics and appliance space. #EdgeAI #VLA #EmbodiedAI #RoboticsHardware #AIWearables #ChinaSourcing #HardwareStartup #PhysicalAI #shenzhenfoundry

While the embodied AI conversation is still dominated by models and simulation, real hardware is already shipping. We visited a factory in China producing multi-camera wearable systems for real-world data capture. The units are in production and the first batch of pre-orders is nearly sold out. The same team supplies camera modules to several leading robotics companies. The factory is running at full speed — their meeting rooms are rarely empty. Photos from our visit below. This is what actual progress on the data infrastructure side looks like. #EmbodiedAI #PhysicalAI #HumanoidRobot #VLA #FirstPersonVision #EgocentricVision

Humanoid robotics is being priced like a hardware story. that may be the wrong frame. The real question is which ecosystem can build the fastest physical AI learning loop: real-world data, VLA models, tactile sensing, supplier depth, manufacturing scale, and deployment feedback. Our view: the winner will not be decided by the best demo video, but by who can turn field failures into better robots faster. Full Research: app.notion.com/p/shenzhenfou…

A small validation from our testing work.🧪 We tested a RK3568 edge setup for live robot-data capture: ⏺️2 camera UDP preview streams ⏺️live IMU calculation ⏺️UART → local print → UDP telemetry ⏺️99.882 Hz average IMU rate ⏺️zero sequence gaps in this run Nothing fancy here, just the kind of plumbing that has to work before robot data becomes useful. video alone is not the whole story. For embodied AI, timing, motion data, and sensor alignment matter too. still early, but a useful step toward cleaner physical data loops. 🧪Full Report: app.notion.com/p/shenzhenfou…