"It looks right" and "it is right" are not the same claim, and most of the painful lessons in this field live in that gap. SC3-Eval sits right on top of it. The pitch is to evaluate robot policies inside a generated video instead of on a physical arm, but generated video drifts. A forward-only model can continue to produce plausible frames long after the action behind them has gone wrong. Looking harder at the picture doesn't help. So they stop asking the model to look forward and make it run the story backward: which action would have caused this footage? When the answer stops matching the command, the rollout has started lying, and you can catch it the moment it does. Covered it in my article for @BotNewsAI #ai #tech

“Back in my day, we had LeRobots in school.” 😂 Some kid is going to say this in 2045 and mean it as ancient history. Children entering classrooms today are getting one of the fastest technological head starts any generation has had: Jetson-powered physical AI, robot learning, simulation, embodied systems. But they also have a strange new problem. Progress is moving fast enough that parts of what they learn can age out before they do. The students in this article reach 18 around 2033–2039. In the span of this story alone, Melania Trump put physical AI into an elementary classroom, Fostering the Future Together expanded to 57 countries, and federal “AI” language became “SI.” By graduation, today’s robotics workflow may look primitive. So I don’t think we prepare them by teaching them to master today’s interface. We teach them to understand the system deeply enough to build beyond whatever gets automated next: electrical systems, hardware repair, physical safety, fleet recovery, deployment architecture, and the infrastructure robots still cannot rebuild for themselves. Fostering the Future Together now spans 57 countries, and the structure around it is becoming recognizable: national representative → technology partner → education system. @FLOTUS and @marcbeckman are helping build the distribution layer. If robotics enters that pipeline, access stops being limited to the schools that can afford to experiment first. Then the harder problem begins: making sure a child educated for 2039 is not being trained for 2026. My article for @BotNewsAI: botnews.ai/news/from-lerobot… @StudyFetch @nvidia @huggingface @intel @MeetMrStewart

Who’s got a home-built robot to show off? Come along tomorrow in SF: luma.com/6rvhcmuc

dimOS v0.0.14 is out 🚀🦾 - A browser UI to control your robot: camera feed, 2D costmap, WASD driving, and a chat box for the agent. - Web SDK for building your own pages. - Zenoh is the default transport now. - A new evals framework to score agents against recordings and live sims. - 3D navigation and ray tracing for dynamic voxel clearing required for global maps. - Record every topic of any blueprint to replay, visualize, or share. - New robots: Galaxea R1 Pro, Deep Robotics M20, Boston Dynamics Spot (experimental), OpenYAM, A1Z. github.com/dimensionalOS/dim…

Getting into your Tesla after leg day

@theworldlabs shipped a world model you can aim. 🎯🌍 Most video models still treat camera movement as language. You describe a crane shot, cross your fingers, and pull the lever again when the drift ruins it. 🤳Atlas, out today from World Labs, takes camera geometry as a native input type. Position, angle, path. Up to one minute of video at 1440p that holds together for the whole run. Under it: a multimodal autoregressive diffusion transformer, pretrained from scratch on text, images, video, and 3D. An LLM builds its context out of tokens in a row. Atlas builds its context out of images pinned to positions in space. Every frame it has seen knows where it stands, so whatever comes next has to agree with the geometry and not just the vibe. 📸Drop two unrelated photos into that spatial context, place them apart in 3D, and the model invents the hallway between them. ➡️ Reconstruction from one to dozens of images, no capture rig, no hundreds of dense views ➡️ Two or three views already give faithful reconstructions, and on sparse-view benchmarks the generalist beats open-source models trained only to reconstruct ➡️ Depth maps are native, so worlds come out as point clouds or Gaussian splats, the same representation already running in Marble ➡️ Third-party human raters picked Atlas over recent video models on camera following in 75 to 94 percent of head to head votes, and the advantage grows as the trajectory gets more complex Then the robotics layer: Two large environments, captured on a cell phone, 24 frames each. Atlas reconstructs the space, then generates the RGB and depth a simulated robot's body-mounted cameras would see as it moves through. The room and the robot's view of the room come from one model, so nothing gets handed off to a second system that can disagree with the reconstruction. Manipulation goes further. From a few casual recordings, Atlas helps build a scene where rigid, articulated, and deformable objects behave, then lets you swap the objects, the lighting, the positions, the motion. Scanning spaces like these traditionally required elaborate and expensive equipment. Now it is a phone in a backpack. Robots were never short on ambition. They were short on rooms to practice in. 🔗 Full technical post: lnkd.in/gJ5UPaQu 🔗 Real-to-sim background: lnkd.in/gZ_FZZwd 🔗 Early access: lnkd.in/gz8YXUMw 🔗 Marble: lnkd.in/g3KSKZ_p Congratulations to @drfeifei, @jcjohnss , @chlassner , @BenMildenhall,@KeunhongP, @ychngji6, @XRarchitect and the World Labs team, and to @YunzhuLiYZ and Changxi Zheng on the robotics side. ⭐ @BotNewsAI botnews.ai for the latest in emerging tech

Rise of the robot pets! 🐕🤖 We’re entering the cutest era of robotics. Engineers are working day and night on the physical AI problems: control, contact, sim-to-real, hardware reliability and getting learned behaviors to survive outside the lab. And somewhere along the way, we get this. @_raghuvamsi filmed one of @eyecandyrobots’ characters wandering around the Presidio, where a little girl spotted it and immediately started laughing and playing with it. There’s something joyfully familiar about that reaction. Before we learn what something is, what it costs or what it’s supposed to do, sometimes we just want to meet it. @eyecandyrobots is building physical AI characters, robots designed around entertainment, personality and interaction rather than industrial work. Founded by @_raghuvamsi, @Mankaran32 and @pr0t0_01, the team brings robotics engineers, animators and designers together to turn animated characters into physical ones. Underneath the cute exterior is serious robotics work: ☑️Animation-based reinforcement learning for character motion ☑️ Artist-centric human-robot interaction ☑️Mechanical, electronics and AI systems designed to scale across different characters ☑️ Simulation, SLAM, deep learning and control-policy deployment on real hardware The transparent little robot really looks like that. I asked Eyecandy engineer @hi_gpaul how they made its clear body: it’s 3D printed using transparent resin, followed by “a bunch of post processing.” Eyecandy is backed by @fdotinc and Lightspeed India, and the team is already thinking beyond a single character. The robots are becoming useful. I’m also very happy that some of them are becoming friends. 🦖 🔗 eyecandyrobotics.com☑️ ⭐️ @BotNewsAI stay connected to emerging tech

If you haven’t heard the Microducks sing you’re missing out 🦆🎶 @huggingface w/ @pollenrobotics just opened pre-orders for Microduck, a 25 cm, under-800 g biped that somehow manages to be both an RL robotics platform and a tiny mechanical duck I desperately want wandering around my workshop. It has 15 degrees of freedom, a camera, two IMUs, an 8×8 time-of-flight depth sensor and an articulated beak. Its policies run onboard at 50 Hz, with skills trained in MuJoCo using reinforcement learning and transferred onto the real robot. But please return your attention to the duck choir. Every Microduck gets its own generated voice, derived from its SoC serial. Put multiple ducks together and they can perform a chorale over Bluetooth. There’s no shared clock. The lowest-ID duck conducts, broadcasts the beat through BLE advertisements, and the others average their timing over roughly 25 beats to stay synchronized. They even work out who sings bass, tenor and the remaining parts without a central controller assigning them. You can also wave your hand in front of the depth sensor and turn the duck into a theremin. Obviously. The software stack is open source, including the SDK, simulator and RL training stack, so owners can retrain behaviors and deploy their own policies. The hardware itself is not open source. Pre-orders opened today at $399, with deliveries targeted before Christmas. My duck is secured 🦆 Microduck: pollen-robotics.com/microduc… Technical docs: github.com/pollen-robotics/m…

This made me miss my grandma fr…. 💔 A robot like this would help her so much

@cerebras just announced the fastest AI server in the world. The CS-4 + Nexus platform reaches 750 PFLOPS, 129.6 PB/s memory bandwidth, 7.2 Tb/s I/O and 2 μs I/O latency, with Cerebras claiming 6× higher system-level performance than CS-3. Sean Lie framed the scaling problem one level higher: frontier inference is no longer just a server problem. It is becoming a rack and cluster problem. For Mixture-of-Experts models, Cerebras can keep the experts on a wafer instead of distributing them across GPUs and repeatedly routing between chips. When traffic does leave the wafer, new direct wafer links and RoCE networking push more bandwidth at lower latency. They’re already mapping GPT-5.6 Sol as a pipeline across wafers, moving activations between them rather than dragging the full model through a web of accelerators. Cerebras projects performance doubling annually, with its 2H 2027 roadmap showing roughly 4× tokens/sec per user and 20× throughput per megawatt. If that curve holds, the next inference race changes shape. The question stops being whether a frontier model can finish the reasoning fast enough. More reasoning passes, agents, verification loops and security checks begin fitting inside the same second and the same power envelope. That compresses the distance between model capability and real-time deployment considerably faster than model benchmarks alone suggest. #cerebras #ai #gpu

We drove these before we had self driving vehicles. Those firefighters didn’t come out of the womb climbing ladders, they had to learn how to walk and gradually progress until they could run. This robot has ran and now it is climbing. Give innovation some grace and it may run with you.

Handroid, at a glance: ✅ An open-source modular robot platform. ✅ One family of electromechanical modules reused to assemble different embodiments. ✳️ The paper demonstrates configurations ranging from dexterous hands to larger articulated robots using the same underlying design philosophy. The engineering preference is notably timely. Instead of treating every embodiment as a fresh mechanical design, Handroid repeatedly leans on the same module architecture and interfaces. Every repeated module is one less custom mechanism to manufacture, stock, simulate, maintain, and support in software. The demo shows a robot changing shape. The paper hints at something intriguing: reducing the number of unique mechanical ideas needed to build one.

I gave a robot a firework. As far as I can find, this is the first publicly documented LeRobot SO-100/SO-101 arm to wield a live firework. Reactive-object manipulation with low-cost robot-learning hardware. The sparkler is the smallest, cheapest object I could find that refuses to stay passive while the robot uses it. Its light changes. Its temperature changes. Its material state changes. Its usable geometry shortens. The functional point moves even when the arm holds perfectly still. That same class of problem appears in welding, soldering, cooking, curing, machining, and other processes where the material continues evolving inside the task. The shared problem is not identical temperature. It is that the robot’s observation, and the state that matters to the task, continue evolving independently of its motor command. A sparkler is a reduced-order proxy, not a miniature welding simulator. It preserves a few hard properties while removing most of the industrial complexity and cost. LeRobot v0.6.0 just shipped GR00T N1.7 and VLA-JEPA. This is step one. The next runs get bigger, and autonomous. #GR00T #LeRobot #HuggingFace #Robot

The transmission inside 1X NEO's new hand shares a family tree with bicycle brakes and surgical endoscopes: tension cables running through compression sheaths, the only routing that carries force across a 3-DoF wrist without caring what the wrist is doing. 1X bet the whole architecture on paying friction tax at a discount: gear ratios of 5:1 to 15:1 where the industry runs past 100:1, low enough that motor current still carries the shape of whatever the fingers are touching. The cinematic footage shows wine glasses and LEGO Duplo but not the force residual at full wrist articulation, six months into a kitchen. We restated the whole field in one accounting standard, priced the parts list on both sides of the Pacific, and traced the tendon loops the launch didn't diagram. Corrections welcomed. #information #robotics #ai #robot #1x #neo

NEO would rather squad up in Fortnite than fold his laundry

The most valuable part of a robot demo may not be the motion anymore. It may be the part of the object it taught the world how to hold. CHORD starts as a dexterous manipulation result and ends as a new accounting system for robot data. On the surface, NVIDIA’s paper reads like another human-video-to-robot-manipulation pipeline: recover the demonstrated contacts, retarget the motion, train the policy. But the reward does not really score the hand. It scores the deal the hand made with the object. A contact point is just an address. It does not tell you whether the finger pinned the object, slid it, tipped it, rotated it, or quietly kept the whole motion from falling apart. The useful thing lives in the contact normal, the moment arm, and the friction cone: the directions that grip can push and twist the object. Pressure never enters the reward, which is the tell. Video cannot reliably recover human squeeze anyway. So CHORD scores wrench support instead. Not “did the robot touch where the human touched?” but “can this grip create the same menu of forces and torques on the object?” That means a human hand and a robot claw can share almost no contact geometry and still count as the same demonstration. Different fingers. Different body. Same object-side authority. Copy the contact positions and articulated objects collapse to 0.000. Score the wrench support and they reach 0.914. That unwelds the demo from the body that recorded it. Once a manipulation demo survives embodiment transfer, it stops being a clip. It becomes inventory: something you can stockpile, rank, license, and rent back to the next policy that needs to know how the world can be held. Somebody is quietly building the object library every future robot policy ends up renting from. Covered it in my article for @BotNewsAI. #ai #tech

Super impressive! Look at the other videos in their blog post 😇

1/🧠Humans are the best robot data source — but video alone misses one thing: force. 2/🙁Tactile gloves capture force — but they're costly and block the real touch manipulation depends on. 3/💪Maybe the future of touch lives on your wrist: surface EMG reads the muscles that cause force — tactile sensing without ever touching a tactile sensor. 4/🔥Want a fully open-source framework — hardware + software — to train your own force-aware learn-from-human-data robot policy? 🚀We introduce ForceBand: Learning Forceful Manipulation with sEMG -- bring force into human videos with sEMG, for force-aware manipulation ⬇️ ✦ Zero-Shot Human-to-Robot Transfer ✦ Force Beyond Vision ✦ Free-Hand Force Sensing ✦ Collect by Anyone, Anytime, Anywhere ✦ Deploy on Any Robot, Any Camera, Any Environment ✦ Open-Source & Low-Cost & Easy-to-Implement Let's squeeze every bit of signal out of human data, and let robots feel the force! 🌐 Website: forceband-emg.github.io/ 📄 Paper: arxiv.org/abs/2606.26093 💻 Code: github.com/Bottle101/ForceBa… 🎥 Video: piped.video/watch?v=Otw6uXZV… 🧵 1/n