Ai2 opens MolmoAct 2 for robots that need to act in the real world

What this is about

Ai2 released MolmoAct 2 on May 5, 2026: an open robotics foundation model intended to improve real-world manipulation tasks. The release includes model weights, an open action tokenizer, and a bimanual tabletop robotics dataset with more than 720 hours of demonstrations.

The timing matters because much of advanced robotics AI remains closed. MolmoAct 2 is not a ready-made home robot. It is a foundation that researchers and companies can inspect, test, and adapt to understand how robots turn images, spatial reasoning, and instructions into physical actions.

What MolmoAct 2 actually does

MolmoAct 2 connects a vision-language model with an action expert. The system observes a scene, reasons about spatial relationships, and outputs movement commands for robot arms. Ai2 describes this family as Action Reasoning Models: the robot should not only imitate patterns but form a useful 3D-like view of the task before acting.

According to Ai2, a single action call takes about 180 milliseconds in the base model and about 790 milliseconds with adaptive depth reasoning. The previous MolmoAct took about 6,700 milliseconds in the cited LIBERO setup. In practice, that is the gap between a robot that visibly pauses and a system moving closer to real-time interaction.

The new Bimanual YAM dataset covers coordinated two-arm tasks such as folding towels, scanning groceries, charging a smartphone, and clearing tables. Ai2 combines it with other robotics datasets so the model is not trained only on one narrow lab setup.

Why it matters

Robots are often discussed as science fiction. In real organizations, the bottlenecks are usually dull and expensive: moving samples in a lab, sorting material, or repeating simple physical routines reliably. Many models fail when lighting, camera angles, or object positions change.

Open releases are valuable because they make progress inspectable. If weights, data, and the pipeline are available, universities, labs, and smaller robotics companies can find real failure modes instead of trusting demo videos. Ai2 reports 87.1 percent average success across several real Franka tasks and 97.2 to 98.1 percent after post-training on LIBERO. Those numbers are not production guarantees, but they provide concrete reference points.

Another point: bimanual robotics is not merely twice as hard as using one arm. The arms can occlude each other, move objects unintentionally, and amplify small mistakes. An open dataset with coordinated tasks is therefore more valuable than a simple collection of single grasping motions. It forces models to reason about order, distance, and collisions.

In addition, the licensing and reproducibility angle matters: when teams can inspect more than an API, they can evaluate dependencies, cost, and safety boundaries more clearly. For Europe and for mid-sized automation companies, that is more practical than another closed robotics demo.

In plain language

Imagine teaching someone to pack a suitcase. A weak system memorizes: trousers left, shoes right. If the suitcase changes, it gets stuck. MolmoAct 2 tries to understand the underlying situation: heavy things go lower, fragile things need protection, and empty space should be used sensibly. That spatial understanding can make actions more robust.

A practical example

Suppose a small lab handles 600 samples per day. A robot needs to remove empty pipette tip boxes and place fresh ones nearby. A rigid script only works if the box, table, and camera position never change. A MolmoAct-like system could infer where the box actually is, choose a grasp point, and after controlled adaptation handle perhaps 80 percent of routine cases. The remaining 20 percent should stay with humans until enough verified data exists for safer deployment.

Scope and limits

• MolmoAct 2 is a research and developer foundation, not a certified safety product for factories, clinics, or homes.
• The reported success rates come from specific benchmarks and lab tasks. Real production environments add dirt, poor lighting, worn grippers, and unclear responsibility.
• Open models lower the barrier to experimentation, but they do not replace robotics safety: emergency stops, force limits, liability rules, and approval processes still matter.

SEO & GEO keywords

MolmoAct 2, Ai2, Allen Institute for AI, robotics foundation model, Action Reasoning Model, bimanual manipulation, open robotics dataset, LIBERO benchmark, Hugging Face, robot AI, vision-language-action