Maxim Lapin @mlap1n

Should we re-train VLA models for every new task? Our answer is NO. Just retrieve. TL;DR: We train a retrieval-conditioned policy once, freeze it, and add new tasks at test time by dropping cheap human-hand demos into a retrieval pool — no per-task fine-tuning.

Today, we enable AutoResearch in the physical world for the first time! Introducing ENPIRE: we give 8 Codex agents a fleet of robots, an allocation of GPUs, and generous token budget. We set them free with a simple goal: solve the task as quickly as possible, keep the robots busy but stay safe, don't waste precious compute. Make no mistake. Then humans step aside and our watch begins. The robot fleet starts to come alive: they learn to look for visual clues, reset the scene, practice novel skills, tinker with control stack, read papers online, debate, reflect, get stuck, and try again directly on the hardware. All we did is to give Codex an API to the world of atoms, and the rest is emergence. ENPIRE is able to solve high-precision tasks like tying zip-ties, organizing fine pins, and installing GPUs all by itself. We also discovered a new type of "physical scaling": 8 robots exploring in parallel improves significantly faster than fewer ones. A part of our NVIDIA GEAR lab now self-improves tirelessly over night. We just read the reports in the morning. /goal: we all take a holiday and Jensen wouldn't even notice ;) We will be open-sourcing everything, so you can host your self-running robot lab at home too! Deep dive in the thread:

With the FIFA World Cup 2026 in the spotlight, see how Boston Dynamics and Hyundai's School of Football collaboration trained the Atlas robot in advanced soccer skills. Blender was used in the pipeline: 80.lv/articles/bosto…

This is wild. NVIDIA just dropped MotionBricks at SIGGRAPH 2026. This AI makes game characters and robots move with 350,000+ motion skills. 15,000 FPS. 2ms latency. 1. Smart locomotion - Characters can now switch movement styles on the fly.

Wow. That's cool. Researchers just released World, an open-source Unreal Engine 5 simulator for training and testing LLM and VLM agents in realistic 3D environments. The platform supports RGB, depth, and segmentation sensors, along with navigation, vehicles, pedestrians, robots and procedural city generation. It's built with a Gym-like Python interface, it allows AI agents to learn physical and social reasoning in complex virtual worlds before real world deployment. We're moving from AI that only understands information to AI that can perceive, reason and act inside realistic simulated environments.

🎮🤖 Introducing MultiWorld: Scalable Multi-Agent Multi-View Video World Models! Current video world models are mostly single-agent, but the real world is multi-agent. We present MultiWorld for action-controllable, multi-view-consistent video world modeling on multi-agent video games and multi-robot collaboration scenarios. Fully open-source — code, models & data! 📄 Paper: huggingface.co/papers/2604.18… 💻 Code: github.com/CIntellifusion… 🌐 Project: multi-world.github.io 🤗 Checkpoint: huggingface.co/Haoyuwu/MultiW… 📚 Data: huggingface.co/datasets/Haoyu…

.@TencentHunyuan HY-World 2.0 has landed,fully OPEN SOURCE ! We kept hearing the same thing: world models look great, but you can't use the output. So we made one you can, real 3D assets you can edit, rearrange, and ship into production. Prompt to engine-ready 3D. For real this time： ✨Directly outputs editable, production-ready 3D assets ✨Seamlessly plugs into existing game pipelines ✨Multimodal input: text, image, videos 🪄Character mode: freely explore streets, buildings, and open worlds with realistic physics, collisions, and no time limits 🔗GitHub: github.com/Tencent-Hunyua…

Matrix-Game 3.0 Real-Time and Streaming Interactive World Model with Long-Horizon Memory paper: huggingface.co/papers/2604.08…

Nvidia released Lyra 2.0 on Hugging Face Explorable Generative 3D Worlds paper: huggingface.co/papers/2604.13… model: huggingface.co/nvidia/Lyra-2.0

Genie3 generates videos. We generate 𝟯𝗗 𝘄𝗼𝗿𝗹𝗱𝘀 you can actually use. Launching tomorrow — Tencent #HYWorld 2.0, an engine-ready World Model🚀 This isn't a video. It's a real 3D scene, all generated & editable. One image in. A whole 3D world out. 🔥Open-source tomorrow

Clone Robotics CEO Dhanush Radhakrishnan: the industry is over-relying on motors and rigid structures. Their own robots use fluid actuation and Myofiber artificial muscles muscles and a full anatomical skeleton to get closer to how real human move.

Most current models rely on reference motions for the humanoid robots to imitate the interactions with objects and environments, limiting their generalization capabilities. We propose LessMimic, an initial step for getting a more general 3D representation for long-horizon humanoid interaction. Using the distance field as a simple but effective representation, we find it promising to extend the current interaction models towards a more generalizable one. We provide a live demo at lessmimic.github.io, where users can customize their interaction objects and use the keyboard to control the humanoid and their interactions!

SimToolReal An Object-Centric Policy for Zero-Shot Dexterous Tool Manipulation paper: huggingface.co/papers/2602.16…

Code2Worlds. language-to-simulation framework. - text to simulation scripts for Blender using Gemini 3; - better scene richness; slashes physics failure; - uses VLM-Motion critic. aigeeksgroup.github.io/Code2Worlds/

I'm excited to introduce ExtremControl — a whole-body control framework that reduces end-to-end latency of teleoperation to 50 ms (just 2.5 control cycles), breaking past the ~200 ms barrier seen in prior systems. Website: owenowl.github.io/extremcontrol/ Code: github.com/UMass-Embodied…

Budapest-based Allonic just raised $7.2 million to develop its so-called 3D tissue braiding system designed to replace traditional robot assembly. The pre-seed funding round, reportedly the largest in Hungary’s history, was led by Visionaries Club, a European early-stage investor group, with other participants including Day One Capital, Prototype, SDAC Ventures, TinyVC, and angel investors affiliated with OpenAI and Hugging Face.

Mike Kalil @mikekalilmfg

4 months ago

A Hungarian startup is generating buzz with its automated manufacturing system for building robots that's straight out of Westworld. Budapest-based Allonic just raised $7.2 million to develop its so-called 3D tissue braiding system designed to replace traditional robot assembly.

3 14 70 17K 19

Meet SceneSmith: An agentic system that generates entire simulation-ready environments from a single text prompt. VLM agents collaborate to build scenes with dozens of objects per room, articulated furniture, and full physics properties. We believe environment generation is no longer the bottleneck for scalable robot training and evaluation in simulation. Website: scenesmith.github.io 👇🧵(1/8)

Here we introduce SAGE: Scalable Agentic 3D Scene Generation for Embodied AI, which can generate sim-ready 3D scenes with agents following user demands at scale, ready for robotic action generation. Paper, code, and SAGE-10k dataset are all released! nvlabs.github.io/sage/

Gemini 3 Flash now uses an agentic "think-act-observe" loop to solve complex visual tasks 🤖 @GoogleDeepMind engineer @ptruiz_dev demonstrates how the model runs Python code automatically to zoom and inspect items, annotate images, and re-visualize data into charts.

LatentLens Revealing Highly Interpretable Visual Tokens in LLMs paper: huggingface.co/papers/2602.00…