Seeing more and more people deploy MolmoAct2 zero-shot on off-the-shelf robots is really rewarding (no fine-tuning, no aligning camera, just work)! We're obviously not there yet, but as a community we're making real strides toward the goal of robots that can simply be deployed right out of the box!

MolmoAct 2 artifacts have been downloaded 400K times in under 1 month. Today we're opening up the full code & training data: everything you need to fine-tune or build on our fully open robotics foundation model. 🧵

MolmoAct2 is now officially available in the LeRobot package 🤩 Try out training MolmoAct2 on your own LeRobot datasets, and deploy it in the real world!

Really blows me away how many people are seeing the power and impact of open-science models like MolmoAct2 being deployed out of the box, without any fine-tuning. That’s exactly the future I envision for robotics foundation models.

MolmoAct2 zero-shot on LeRobot SO101 arms even in a cluttered real-world setup, it nails the grasp → drop into bowl task with only a front and side camera!! No fine-tuning, fully open-source action model going straight from reasoning to manipulation. This is the future of open source robotic models!!! @DJiafei @hq_fang @allen_ai

MolmoAct2 is also fully integrated into LeRobot! Have fun building, deploying and experiencing robotics generalist model being deployed out of the box!

MolmoAct2 is landing in LeRobot! Ai2's open Action Reasoning Model combines a Molmo2-ER vision-language backbone with a flow-matching continuous action expert to predict robot action chunks from images, language instructions, and proprioceptive state. An open robot foundation model built for real-world control, with strong out-of-the-box performance and easy fine-tuning in LeRobot. Pick-and-place inference running on NVIDIA DGX Spark! Blog: allenai.org/blog/molmoact2 Paper: arxiv.org/abs/2605.02881 Thanks to @allen_ai @DJiafei @hq_fang

MolmoAct2 is most ideal for post training RL! Excited to see more building on it

Molmoact2 fine tuned on 35 examples running inference with RTC. Trained the expert head for 20k steps, batch size 8. Very happy with this result and it looks like a great place to build on for experimenting with RECAP, RLT and other methods

Happy to see MolmoAct2 being deployed so easily!

happy to show what @livekit can help bring to life over the next few months, we’ll stronghold our position further as the go-to for real-time robot deployment and interaction this is just a glimpse of what we can do btw, the thread on the implementation of this demo is here x.com/pham_blnh/status/20577…

Got the i2RT YAM bot running with MolmoAct2 today. Off-the-shelf capabilities continue to impress, even in terrible 2am lighting. Prompt: "place all objects in container" Waiting on leader arm parts to try fine-tuning @allen_ai @DJiafei @i2rt_robotics

One week since we released MolmoAct2 and @irenekarrot already had it running on her brand new YAMs out of the box. Robotics is supposed to be hard. Models that just work feel like magic. 🤖

A week since MolmoAct2 (@allen_ai) launched, and the community response has been incredible — deployments, builds, and out-of-the-box rollouts all over the place. 🤖 We've now released the fine-tuning code, with @LeRobotHF support coming soon. github.com/allenai/molmoact2 Below: a 2x rollout fine-tuned on just 50 demos 👇

A side quest I've been doing in my unemployment era is DMing authors of recent robotics papers and asking them to open-source their evaluation episodes onto HuggingFace. Thanks to @DJiafei for open-sourcing the evals that went into Molmoact2, this is a great contribution to the field. If you are robotics person and want your evals shared data to be signal-boosted on X, let me know huggingface.co/collections/J…

Open science is in our DNA @allen_ai and i believe openness is key to moving the field forward as a whole.

Had a chance to fully read the MolmoACT2 paper today. Imo, the ablation results are the most exciting part. So many ideas popping.

I keep getting asked: why are we still evaluating VLAs on LIBERO? My take: it's adoption. So many people benchmark on LIBERO that the community follows suit. Yes, 98.2% vs 98.7% is hard to distinguish, but it still tells us something about policy efficiency, at least for quick fine-tuning on that domain and those embodiments. Better benchmarks have emerged since. MolmoSpace, for instance, evaluates more than just VLAs. And even though MolmoAct2 currently tops it among VLAs, there's still plenty of headroom for other approaches like TipTOP (Open-world FM TAMP) to push past it. The ceiling is nowhere near saturated, which makes it great for evaluating at scale in simulation. Robotics evaluation is far from solved, and no evaluation is perfect yet. But that's exactly what makes the field exciting with so many open questions still on the table.

Just a few days left to enter our Embodied Reasoning in Action workshop challenge @CVPR ! Submit your entry by May 23, 2026 to be considered. You can take part in either challenge or both for a chance at winning cash prizes up to $6,000 USD, sponsored by @BitRobotNetwork, @NVIDIARobotics, and @cortexairobot. Winners also get the opportunity to present their solution in person at the conference. 🤖

🎙️The AI Talks Building Open Science Robotics Foundation Models - by Jiafei Duan @DJiafei & Haoquan Fang @hq_fang What is critical for real-world robotics? 🤖 Explore MolmoAct & MolmoAct2 - reliable deployment via Action Reasoning, adaptive-depth control, and open datasets. ⬇️