谁是藏镜人

Curiosity takes us to unexpected places; this time we gave Atlas wheels. At its core, high-performance RL is about controlling the dynamics of two interacting loops: the policy execution loop nested inside the policy learning loop. That’s why some of the best RL engineers start as control experts. Here, a more detailed explanation from Andy Wu Blog: bostondynamics.com/blog/atla… Video: youtu.be/NL65Awuv8gU

How about: Whole-body motion tracking for the Unitree G1 in under a minute :) XMimic trains one policy on all 40 LAFAN1 clips in 54 seconds on 4 H100s. 3.5x faster than BlinkMimic and 85x faster than SuperMimic. Code: github.com/nico-bohlinger/xm… Report: nico-bohlinger.github.io/pap…

@elonmusk on the hardest part of building Optimus: "The fine motor skills are incredibly difficult to get right. Like shockingly difficult." Signing your name, or driving a self-tapping screw into wood, takes precision and force at the same time. "There's no hand design that I've seen, not even sort of a demo that does that." "But that's what we're trying to achieve with Optimus." @Tesla_Optimus @theXtakeover

Successful test flight of our big 200kg open-source drone: Caribou 👏 github.com/Arrow-air/project…

What if a Gaussian splat scene went full Matrix? 🟩 New @playcanvas shader effect: digital rain for 3D Gaussian splats. Some splats turn into camera-facing glyphs lit by falling drops. The rest thin out, shrink and tint. Open source, one script, runs on WebGL2 and WebGPU.

20 days ago we built a system that writes GPU code. Today its kernels outperform the state of the art. We just open-sourced 9 megakernels, which outperform NVIDIA’s stack on DeepSeek and run self-driving, robotics, protein folding and world models 2–3× faster than the current best systems. Every model runs in a single launch instead of thousands, and most of our kernels were generated in under an hour.🧵

Meet MIRRAX-150: The #Robot Snake Built for Nuclear Rescue Missions by @lukas_m_ziegler #AI #Robotics #EmergingTech #Engineering #Innovation #Technology

A finer touch for humanoid robots. 🤖 At #CoRL2026, NVIDIA is introducing GR00T 2, an open world action model for humanoids, alongside new open industrial benchmarks from @RLWRLD_ai and @UCBerkeley built on Isaac Lab-Arena. Explore how NVIDIA’s open world action models, scalable simulation tools and agentic workflows help researchers train, test and refine robot skills, from connecting cables to assembling small components. Discover more ➡️ nvda.ws/4ixvdZV

This robot can pick up a sock with one button that its makers describe as “beak to the floor.” The move itself is simple: the robot crouches, closes its beak around the object and stands back up. A separate neural network triggers it, and it exists even in the browser version of the simulator, where the G key calls it, in legs mode only. That simulator deserves its own paragraph. It runs entirely in the browser, with no server: the physics is compiled to WebAssembly. Other visitors show up next to you as translucent ghost ducks, up to three at a time. They connect directly over WebRTC and send their positions ten times a second through public Nostr relays. The modes there are: running at a set speed, sitting and standing, a roll, and a kick with either leg. The kick is “blind”: a one-shot action in a 0.5 s window. When wheels are fitted, a different network takes over: driving on four passive wheels at a top speed of 0.6 m/s, and a crouch-glide mode that lowers the robot over about 3.5 s before it stands up. The main repository is written in Rust in a single Cargo workspace, with no framework. At the time of the project page it had 13.1k stars and 1.6k forks, and its root holds files called AGENTS.md and CLAUDE.md, instructions for AI agents that write code. Pollen itself warns that this 25 cm robot, about 14 cm wide, should not be mistaken for a household robot. The README ends with the line: “No duck was harmed in the making of this robot. Several were consulted.”

Can a world-action model learn to predict touch without any tactile pretraining? We introduce VTAM, an efficient visuo-tactile world-action model. With ~100 demos per task and no tactile or action pretraining, it is comparable to tactile models pretrained on 30K+ hours of visuo-tactile robot data, and is 6× faster at inference. 🧵 📄Paper: arxiv.org/abs/2603.23481 🌐Web: plan-lab.github.io/projects/… Code and data are open source! 💻 Code: github.com/PLAN-Lab/VTAM 🤗 Data (xArm + GelSight, 3 tasks): huggingface.co/datasets/Jens…