This robot arm picks with bat sonar! 🦇 When a bat echolocates, the tiny difference in timing and volume between its two ears gives it direction from a single ping. Recreating that on a robot normally means several receivers and amplifiers, which gets expensive fast. This build by @karolina_dubiel goes the other way. One HC-SR04, the $3 ultrasonic sensor from every Arduino starter kit, firing a 40 kHz burst and returning a single number: distance. How do you get direction out of a sensor that only measures distance? You move the sensor. A scanning routine sweeps the arm across a grid, pings at each spot, and tags every reading with the current joint angles. One cheap sensor plus known geometry rebuilds a height map of the table. Then it becomes a precision problem. Under gravity the shoulder and elbow kept stopping short, smearing the map by up to 13 mm. The fix: tighten the joint tolerance from 0.9° to 0.26°, and re-command each joint with its measured shortfall, up to six tries, until it actually lands. And because the beam is a ~15° cone, edges blur into neighboring cells, so a classifier only accepts a 5×5 cm cell when readings agree across multiple sweeps and the median height matches the object. That’s just brilliant. 😮‍💨 🔗 Here's the project: Be like Karolina, build, build... and build! 🦾 ~~  ♻️ Join the weekly robotics newsletter, and never miss any news →

🚨 BREAKING: Meta reportedly has built a prototype humanoid robot, in partnership with Chinese startup @FourierRobots. Fourier is one of China’s most capable humanoid hardware companies, sophisticated bipedal platforms, strong actuator supply chains, deep manufacturing expertise. I've met them couple of times, and was really impressed with the robots. Their roots are in exoskeletons + rehab robotics. The robot is called Max, and here’s what we know about it: → Built in approximately three months → Weighs around 106 pounds (48 kg) → Meta is already building a second version → Meta has no plans to sell these units → Framed internally as the physical embodiment of Meta’s Muse agent (@alexandr_wang bring it on!), designed for household tasks For Meta, partnering with a Chinese hardware company to build a prototype is the fastest path to having something physical to work with while its AI and software teams develop the intelligence layer. This isn’t a manufacturing robot, but rather home robot. Similar move to acquisition of Fauna Robotics by Amazon earlier this year. Meta is building a physical body for its AI agent, the same way a phone is a physical body for an operating system. Every major AI lab (Sam Altman also announced that will reveal a humanoid robot in 2027) is arriving at the same conclusion. The next frontier is not software on screens. It’s AI in the physical world. It was first reported by @theinformation (@jyoti_mann1). Watch this space very closely! 👀 ~~  ♻️ Join the weekly robotics newsletter, and never miss any news →

What is Foxconn's approach to physical? 🤔 During HHTD in Taipei I sat down with Leo Guo , who runs physical AI and robotics at Foxconn, and had a chance to discuss with him what's their strategy. So first of all @HonHai_Foxconn isn’t building its own robot brain from scratch. Right now half the robotics world is racing to train its own foundation model. They don’t build the foundation model. They partner with @NVIDIARobotics and run on GR00T 2.0. Foxconn’s edge isn’t the model, it’s the factories. They have a HUUGE range of real industrial scenarios, so they collect real production data and use it to post-train the model on tasks that actually happen on a line. That data goes back to NVIDIA to make GR00T better, which comes back stronger, which then collects even better data. A flywheel, spinning on an asset almost nobody else has at this scale. The hardware follows the same logic. Build in-house where they can win, pick the best partner where they can’t. → No in-house foundation model, GR00T 2.0 via NVIDIA instead with own data stack → Real industrial data from Foxconn factories feeds the post-training → A data flywheel: deploy, collect, post-train, repeat → Components built in-house only when they beat the market, like motors Their deployment scale pulls genuinely competitive pricing from suppliers The result is a robotics stack that’s cost-competitive in a way very few can match, built from the best parts on the market and their own data. Build what you’re best at, partner for the rest, and let the factories do the talking. Thanks for having me Foxconn! 🦾 ~~ ♻️ Join the weekly robotics newsletter, and never miss any news →

Foxconn is mass-deploying humanoids that are assembling NVIDIAs super computers Rubin Veras with 99% efficiency If you want to deploy humanoids, you must have a customer who’s is incentivised to use robots. Focus on these who are really willing to go hand-in-hand into the deployment journey.

Robotics at University of Michigan is on 🔥! Plus, it's completely open-sourced and free. University of Michigan Robotics Department put their entire robotics program on GitHub. Everything. Code. Textbooks. Examples. Simulations. That's what inside: → ROB 101: Computational Linear Algebra. Foundation course on the math that actually powers robots. → ROB 201: Calculus for the Modern Engineer. They wrote a textbook from scratch (GPL-licensed, on GitHub) because standard calculus doesn't address what engineers need. → ROB 204: Introduction to Human-Robot Systems. How humans and robots actually interact. → ROB 311: How to Build Robots and Make Them Move. The gap between theory and practice. → ROB 501: Mathematics for Robotics. Graduate-level rigor with code examples and interactive Jupyter notebooks. If you want to learn robotics math from one of the world's best robotics departments, it's free. Over 2,800 GitHub stars on the math courses alone. Thousands of people learning robotics, just because it's open for everyone :) This is how you build the next generation of roboticists! 🔗 Access it here: ~~ ♻️ Join the weekly robotics newsletter, and never miss any news →

Wait… these glasses send radar through your face. 📡 You silently mouth a word, your speech muscles twitch a fraction of a millimeter, the radar reads it, and it’s text at 200+ wpm. No voice or typing. You just mouth it. Awesome use of sensing hardware :)

The world calls Foxconn a “box mover.” 📦 Their Chairman, Young Liu disagrees. Today I met Young Liu, Chairman of @HonHai_Foxconn, here in Taiwan, and that was a conversation that will stay with me for a long time. First of all, you don't meet a chairman of one of the biggest companies in the world every week. Alright, so here's the reaaally interesting part! He said the world still tends to see Foxconn as a “box mover”, a company that just assembles electronics someone else designed. Then he brought his numbers: → Over 1 million employees worldwide → 120,000+ engineers → 78,000 system designers → ~2,500 patents in 2025 alone That’s not a workforce built for moving boxes. Definitely not. That’s a workforce built for designing what goes in them. And that’s the shift he’s driving. Foxconn is not doing the vertical aggregation but vertical integration! → Vertical aggregation: assembly, molding, producing the parts (just being the contractor) → Vertical integration: co-design, co-development, co-validation, participating in the full end-to-end process You see it in the AI server business, where Foxconn now works hand in hand with NVIDIA, and where AI servers are set to make up more than half of its server revenue this year. And the part that landed closest to home: Poland. 🇵🇱 Foxconn is building toward EV production in Poland with ElectroMobility Poland, a technology hub in Jaworzno, platforms from its Foxtron arm, and first cars targeted for 2029 under Polish control. Actually while asked why Poland, he said it reminds him of Taiwan. He pointed to the flexibility, the dynamism, and the work ethic. The company everyone files under “assembly” is quietly becoming a design and engineering partner, on three continents at once. Damn, what a day! Expect tons of content from Taiwan! 🇹🇼 ~~  ♻️ Join the weekly robotics newsletter, and never miss any news →

A quadruped just solved a Rubik’s cube, with its feet! 🐾 A four-legged robot that normally just walks, autonomously solving a giant Rubik’s cube using only the two feet it also stands on. This is new work from ETH Zürich and NVIDIA Robotics. They call it pedipulation, manipulation with feet. The robot walks around the cube and scans it with an onboard depth camera. SAM 3 segments the cube, ICP locks in its 3D pose, and the facelet colors get classified to rebuild the full cube state. A standard solver plans the moves. Then learned skills execute them. The hardest skill is turning a layer: one foot pins the cube in place while the other pushes the top layers 90°. Pin and push at once, on legs that are also holding the robot up. → Trained in simulation with RL, then distilled into policies running on just depth + the robot’s own joint sensing → ~2 cm pose error and under 3° rotation error out to 2 m → Facelet color read with 100% accuracy indoors, 98.7% outdoors → Three full autonomous solves on real hardware, ~30 steps each, 8–9 minutes A party trick on the surface. Underneath, it’s a quadruped completing a long, precise, multi-step physical task with nothing but its feet and its own eyes. That’s a real step toward legged robots that do more than walk. 😮‍💨 🔗 Project page here: ~~  ♻️ Join the weekly robotics newsletter, and never miss any news →

Here we go Taiwan! 🇹🇼 Hon Hai Tech Days is kicking off in Taipei, hosted by @HonHai_Foxconn. I just sat down with Leo Guo, their GM of Robotics, who walked me through Foxconn’s robotics program and where they’re currently putting their focus. I really liked their deployment approach. So, they build their own hardware (robot arms, humanoids) and their own AI models, but they also deploy third-party robots and layer their own software on top. They experiment a lot, look for new use cases, and are very curious. Looking forward to learn more here on-site! Stay tuned for more content! 😎 ~~ ♻️ Join the weekly robotics newsletter, and never miss any news →

JUST IN: @nebiusai and @positronic_ro just launched a Physical AI Competition, and the prize pool is $20K+ in GPU and evaluation credits. On the DROID: Everyday Tasks leaderboard at over 1,000 episodes, 100+ common tasks, all on the same Franka arm — the best open models complete only ~70% of tasks on average. The weakest: ~12%. Robotic manipulation is far from solved. And the gap between models is wide enough for a small team with a good idea to make a real difference. That’s the whole point of this competition. → Open to all teams — register at → First 50 teams to submit in week one get $100 in Nebius GPU credits → Models evaluated first in simulation → Top 10 advance to testing on a real physical robot arm → Competition runs October to December 2026 → Powered by Nebius GPU infrastructure and Physical AI Workbench The historical parallel the organisers draw is apt. In the 2004 DARPA Grand Challenge, no autonomous vehicle finished the course. A year later, five did. Open competitions with shared rules create turning points. Robotic manipulation needs its DARPA moment. 🔗 Register here: ~~ ♻️ Join the weekly robotics newsletter, and never miss any news →

🚨 BREAKING: @BleuRobotics01 has raised €5M pre-seed to solve the humanoid robot deployments. Co-led by Contrarian Ventures and OVNI Capital, with participation from Kvanted, WEPA and Charlie Songhurst, technology investor and Meta board member, and I'm also an angel Funny story, I found Benoît's company 3 days after he incorporated it, through French company registry last December, real angel grind haha. Of the 20,000+ humanoids produced last year, only 10% reached real-world applications. Most robots today rely on end-to-end models that require hundreds of hours of training data and months of deployment time. For manufacturers, that’s prohibitive. Bleu’s goal is to reduce the full humanoid deployment process to under one hour. Instead of training on massive datasets, the model infers the most likely intent behind an operator’s demonstration, the task’s goal, its constraints, how to reproduce it reliably. A robot learns a new industrial task from as few as ten demonstrations. → 10 demonstrations to learn a new task → 99%+ accuracy required for production lines, and delivered → Up to 100x more data-efficient than state-of-the-art alternatives → Plugs into any humanoid platform → Operators teach tasks directly at the workstation, no coding required → Robot can be retrained as production needs change or moved to another site The founding team is the right one for this problem. Benoît Berkoukchi managed large assembly line teams at Renault. Jean-Baptiste Mouret and Serena Ivaldi are research directors at INRIA, 15 years at the frontier of AI and robotics, published in Nature, organisers of leading humanoid robotics conferences. Industrial expertise. World-class research. And a software-first approach to a hardware-dominated industry. First deployments already underway at major automotive manufacturers and WEPA, one of Europe’s largest hygiene paper manufacturers. Europe is not waiting. 🇫🇷🤖 ~~ ♻️ Join the weekly robotics newsletter, and never miss any news →

Christoph Schell runs KUKA, the company that built one of the first industrial robots, Famulus, in 1973 and now has close to 600,000 robots working in factories around the world. Now KUKA is betting on physical AI. The idea is to combine the deterministic automation KUKA is known for with robots that understand a situation and act on intent, so that even older machines can learn new tricks. I sat down with Christoph to talk about where the training data for physical AI really comes from, why KUKA is growing faster than the market in China, robotics as a service, and why he sees humanoids as one form factor among many mobile robots rather than a category of their own. Christoph told me why this shift feels different from anything he’s seen before: “We’re not disrupting the old industry. We are evolving it.” This is a conversation about physical AI on real factory floors, the business models that come with it, and why the industry’s deployment timelines may be too ambitious. 00:00 Intro and KUKA’s first robot 00:40 What physical AI changes for existing robots 01:45 KUKA’s installed base and factory data 02:40 Automation 2.0: the bin-picking example 03:35 Selling intent-based automation to factories 04:45 Where physical-AI training data comes from 06:15 What data gets you to the 99,99999% reliability 07:10 Robotics as a service 08:25 KUKA’s growth in China 09:40 What does scaling mean for KUKA? 10:45 Why KUKA built a US AI team? 11:35 Is the robotics timeline too ambitious? 12:53 Humanoids versus other mobile robots 14:35 What industrial leaders will need in 2036 ~~  ♻️ Join the weekly robotics newsletter, and never miss any news →

Qualia is now scoring very high on three important public robot-data benchmarks, scored exactly the way NVIDIA scores them. (Our setup reproduces NVIDIA’s numbers for GPT-6 Astra and Flora within about a point. Side note flora is surprisingly good) Cutting robot video into steps, describing them, ranking recordings, judging how far a robot got: Qualia is ahead of Flora and GPT-6 Astra on every one. We are just getting started, and expect Qualia to improve a lot more in the coming days. and weeks. Hope to get more of you to try this soon.

Open source prosthetics! 🦿 A powered prosthetic leg runs its control loop at a few hundred hertz, on a battery computer strapped to someone’s shin, and the failure mode is literally a person falling over. The Neurobionics Lab at Michigan says it plainly about their Open Source Leg: labs each build their own prosthesis, the mechanical designs differ so much that control strategies can’t be fairly compared, and the field’s impact on amputees suffers for it. So they built one leg for everyone to use. → pancake motors from the drone industry, geared down to trade speed for torque and finer control → belt drives, no precision-machined parts, easy to assemble and take apart → modular: knee, ankle or both, with onboard power and control electronics → optional series elasticity at the knee, so researchers choose the stiffness → a Raspberry Pi running the control Elliott Rouse, who directs the lab: “we hope our bionic leg will unite researchers with a common hardware platform. AWESOME STUFF! ~~ ♻️ Join the weekly robotics newsletter, and never miss any news →

The fastest company ever to hit $100m ARR! 💸 It's coming from Europe, and it's in the physical AI space. During the NVIDIA GTC I had a chance to speak with Artjem Weissbeck from @microagi to understand what sits underneath that. The mission hasn’t changed at all. Namely, deploy a billion robots! 🤯 They started in data collection. Now it’s models, hardware, data, and the research and people around all of it. End to end. Most of physical AI is specialising right now: a model company, a data company, a simulation company, a hardware company. As we discussed their focus is large manufacturers who see robot deployment as how they make it into the next decade. A manufacturer is not buying components to assemble. They’re buying a working deployment, with safety and accuracy someone has to sign off on. When the buyer is a plant or a warehouse, end to end is the shape the product has to take. It also explains the sequence. Starting in data and climbing the stack teaches you precisely what’s missing, because you are the one who keeps hitting it. @bercankilic, @heimmoritz, @ZenoInMotion, @artjemw it was awesome to catchup!! @NVIDIARobotics ~~ ♻️ Join the weekly robotics newsletter, and never miss any news →

Robots at Dexory are being continuously tested 🔁 Over and over again :-) I asked Andrei Danescu about testing, and that was the answer. Which is the thing simulation cannot touch. You can simulate perception, dynamics, contact, a whole snowstorm. You cannot simulate twelve months of bearing wear, cable flex, thermal cycling and dust. Reliability data arrives at one second per second. Here's the YouTube video from my visit in their factory (subscribe if you like robots hehe):