AI2 unveils MolmoAct 2

AI2 has released MolmoAct 2, an open foundation model for real-world robots, together with the MolmoAct 2-Bimanual YAM dataset. It released the model, data, and an adaptive reasoning approach together, and plans to release the training code soon. The new model is designed to handle multiple manipulation tasks without task-specific fine-tuning, and is up to 37x faster than the previous generation.

The core architecture is a KV-cache bridge connecting a Molmo 2-ER backbone to an expert that generates robot actions via flow matching. It was further trained on about 3 million embodied-reasoning examples to strengthen pointing, object detection, multi-image reasoning, spatial reasoning, and image/video-based question answering, achieving an average of 63.8 points across 13 benchmarks, surpassing GPT-5, Gemini 2.5 Pro, Qwen3-VL-8B, and GR-ER 1.5. The MolmoAct 2-FAST Tokenizer was also open-sourced as an in-house reimplementation and released alongside it.

Speed was also greatly improved. A single action call takes about 180ms for the base model, and about 790ms when using MolmoAct 2-Think with adaptive depth reasoning, both far faster than the previous MolmoAct's 6,700ms. Depth tokens are predicted only in the regions where they're needed, yielding 17% faster speed compared to predicting all depth tokens, while maintaining efficiency even as 3D spatial reasoning goes deeper.

Training centered on MolmoAct 2-Bimanual YAM, currently the largest publicly available bimanual tabletop manipulation robot dataset. It contains over 720 hours of two-arm collaborative demonstrations, more than 30x the robot data used in MolmoAct. It blends SO-100/SO-101, filtered DROID Franka, Google Robot BC-Z, Fractal, Bridge WidowX, and the original MolmoAct data, absorbing a variety of arms, cameras, and control schemes. Relabeling also increased the number of unique labels from about 71K to 146K.

Evaluation results were also clear.

• MolmoBot: an average success rate of 20.6%, ahead of π0.5's 10.3%.
• RoboEval: 0.443, higher than π0.5's 0.405.
• Franka zero-shot: an average success rate of 87.1%, well above MolmoBot (48.4%) and π0.5 (45.2%).
• LIBERO post-training: 97.2%, with MolmoAct 2-Think reaching 98.1%.
• Cortex AI external evaluation: an average of 0.51, outperforming OpenVLA-OFT (0.36), π0.5 (0.32), Cosmos Policy (0.16), and X-VLA (0.05).

At deployment, natural language instructions and visual traces can be used together to guide robot trajectories, and reference hardware was also released, including two YAM arms, an Intel RealSense D435, two D405 cameras, an extendable mount, and a simple tabletop workspace. In a pilot at Stanford School of Medicine's Cong Lab self-driving wetlab, it showed promise by handling routines such as sample transfer and equipment manipulation in CRISPR experiments, and internal stress tests were also conducted on instruction rephrasing, object position changes, distractors, and object substitution. However, limitations remain, such as gripper occlusion, control latency, fine manipulation, and depth errors in 2D visual traces.