Robotics has spent years scaling demonstration data and model size. What happens when we scale EXPLORATION for robot agents? We're incredibly excited to share RAX, an agentic robotics system that learns by scaling online exploration. A coding agent connects directly to the robot’s perception and control APIs, learns from feedback, and builds programs that perceive, plan, and act. The model stays frozen. As exploration scales, the agent improves its programs and builds up reusable skills. The results: • ZERO teleoperation or action demonstrations on real robots—from insertion with 0.1 mm clearance and air-fryer manipulation to deformable objects and long-horizon mobile whole-body control. • +37 percentage points on average over leading learned policies across five simulation benchmarks: RoboDojo, BEHAVIOR-1K, HumanoidArena, DexJoCo’s bimanual tasks, and LIBERO-Pro. • One framework spanning single-arm, bimanual, dexterous, humanoid, and mobile robots, with new embodiments connected through APIs. The craziest part: on each real-world embodiment, going from scratch to mastering the tasks took roughly two hours. No teleop. No model finetuning. We’re open-sourcing the full framework AND every experiment session. We can’t wait to see what people build with it! Project + sessions: Paper: GitHub: N Project fully led by @RunhanH and @xczhangxx who are absolutely awesome. How it works, experiments, and robot videos below 🧵 (1/N)

ASPIRE can now solve a scrambled Rubik's Cube with a real robot arm 🧩🦾 With current frontier models, it reaches 99–100% success on LIBERO-Pro, Robosuite and BEHAVIOR-1K tasks. It's open source. Clone it and run it on your own machine:

Super excited to share some of the work I’ve been doing at the GEAR lab @NVIDIAAI. After three months, we’re reintroducing ASPIRE: Agentic Skills Discovery for Robotics (v1.0). Enough of Astra ✨ doing pick-and-place—let it learn reusable strategies and solve a Rubik’s Cube online. Quick 🧵 for details.

We are open-sourcing V1.0 of our agentic robotics framework, ASPIRE, with full real robot skill discovery and self improvement capabilities! On the day Astra arrived, our lab switched the backend model and saturated all the benchmarks to ~99% over night. And the frontier is already on the move: - Opus 5.5 slightly edges out Astra on our robot simulation bench. - Sonnet 5.5 hits the Pareto frontier of cost vs. robot capability, $12/task. We first introduced ASPIRE and /robot skill library in June, way before the big models could reliably do robotics. A rising frontier tide lifts all boats, but Astra was a 0->1 tsunami on many tasks. Now you can also feel the Physical Bitter Lesson at home like we did on that fateful day ;) links in thread

ASPIRE is now open-source! Since our initial release, we've made exciting progress: 🤖 Real-world robot manipulation with GPT-6 Astra. 🔄 Continual skill refinement on physical robots. 📈 Saturated simulation benchmarks: 99% on LIBERO-Pro (80 tasks), 100% on Robosuite (7 representative tasks), and 100% on BEHAVIOR-1K (2 representative tasks). Code: Project: 🧵 More results in the thread below.

We’ve all seen the cool “Astra for robotics” demos. But behind the scenes? Slow rollouts, hand-engineered APIs, and/or simulated behaviors that struggle on real robots. How do we make coding agent actually work in the physical world? 🦾 Introducing SimEX: Simulation-Integrated Robotics AutoResearch 🧵

What is the role of simulation in the era of Astra? Traditionally, in sim2real transfer, we transfer data or policies, which requires relatively accurate simulation. What excites me about “agentic sim2real” is that we can also transfer tools, knowledge, and procedures. A roughly correct simulator can give agents a place to test their physical intuition, run experiments, and build reusable capabilities. That’s the idea behind SimEX, led by @JiahengHu1: develop capabilities through simulated experimentation, then use a few real-world trials to refine both the tools and the simulator.

Practice makes perfect. A truly agentic robot shouldn’t just know what to do. It should also figure out how to get better. What if a robot could watch a demonstration, figure out what it needs to practice, and learn how to get better through practice? Introducing RPG: Guided Self-Improvement for Embodied Agents, a real-to-sim-to-real pipeline that turns real-world demonstrations into targeted practice. RPG reconstructs practice tasks from demonstrations, uses demo videos to guide skill improvement, and continually updates a shared skill system from failures. On real robots, RPG achieves 30/30 successes across 3 tasks. In simulation, it reaches 95.0% success across 22 manipulation tasks after 15 rounds of self-improvement. Demo → Reconstruct → Guided Practice → Self-Improve → Go Real

We have seen quite a few capable agents controlling robots. But an important next question is: how can we build an automatic pipeline for these agents to keep improving themselves? Check out RPG. Starting from an offline robot dataset, RPG automatically figures out what capabilities need to be practiced, builds sim environments for those skills (practice tasks, rather than digital twins), and then improves the robot’s skill library through repeated rounds of practice and failure analysis. Then, the improved system also transfers back to the real world, achieving 30/30 successful hardware trials across three tasks. Today we use datasets mainly to train robots. In the future, datasets can also tell robots what they should practice next. As robot data becomes increasingly abundant, automatically turning those experiences into targeted practice and continual self-improvement feels is really promising!

Very excited to share LoGo, my internship project at @theworldlabs on post-training world models! We found that reward design matters a great deal in post-training long-horizon video gen for 3D consistency, echoing what we see in other domains, e.g. LLM reasoning. Check out more 👉

Super excited to share Dyna-2, the first robot foundation model trained on over 1 million hours of human data. At this scale, we saw the emergence of a cross-embodiment transfer scaling law: training on increasing amount of human data not only improves model prediction on held-out human data but also on robot data the model has never seen before 🤯 We validate that this transfer scaling law does translate to on-robot performance across 3 different robot platforms, and uncovers that both training objective and data matter greatly for this emergence. Crucially, our results position video as a new scaling axis for physical AI. Beyond these scaling-law oriented results, we also dissect Dyna-2 greatly, going in-depth on its many capabilities, including language following via world modeling, enhanced robustness and precision, zero-shot customer site deployment, and some cool video generation results! Read our technical blog post for more details! I do think this is a very important result that shows a very different path for robot foundation models from the ones we are marching on. Really excited for the road ahead. We have more releases coming, stay tuned!

Very cool to see @netflix releasing video datasets and models on @huggingface. They have a strong video AI team and would be amazing to see them open-source more!

Thrilled to see Vera, my internship project at @netflix, shared here! Vera brings editable layers to diffusion models for controllable video editing. Check out our project page for more details, the paper, and the dataset: