allenai / Molmo2-ER

Molmo2-ER

Molmo2-ER (Embodied Reasoning) is a 4B vision–language model specialized for the embodied perception skills that downstream action models depend on: scene understanding, pixel-accurate pointing, multi-image and egocentric–exocentric correspondence, and video temporal reasoning.

It is built on top of Molmo2 (Qwen3-4B backbone + SigLIP2 vision encoder) and serves as the vision–language backbone of the MolmoAct2 action reasoning model.

Highlights

• Outperforms every open-weight baseline as well as the strongest closed-source models — including Gemini Robot-ER 1.5 Thinking and GPT-5 — on 9 of 13 established embodied reasoning benchmarks (Point-Bench, RefSpatial, BLINK, CV-Bench, ERQA, EmbSpatial, MindCube, SAT, VSI-Bench).
• Overall average 63.8% , a +17 point improvement over the Molmo2 starting point.

Training

Molmo2-ER is trained from the released Molmo2 checkpoint with a two-stage specialize-then-rehearse recipe:

All other hyperparameters follow Molmo2 .

Resources

• Code : https://github.com/allenai/molmo2
• Base model : Molmo2-4B

Usage

See https://github.com/allenai/molmo2 for inference, evaluation, and training code.

License

Apache-2.0.

Model and Hardware Safety

MolmoAct2 generate robot actions from visual observations and language instructions, but their behavior may vary across embodiments, environments, and hardware configurations. Users should carefully validate model outputs before deployment, especially when operating physical robots or other actuated systems. Where possible, actions should be monitored through interpretable intermediate outputs (adaptive depth map), simulation rollouts, action limits, or other safety checks before execution on hardware. The model’s action space should be bounded by the training data, robot controller limits, and task-specific safety constraints, including limits on speed, workspace, torque, and contact force. Users should follow the hardware manufacturer’s safety guidelines, use appropriate emergency-stop mechanisms, and operate the system only in a safely configured environment with human supervision.

Citation

@misc{fang2026molmoact2actionreasoningmodels,
      title={MolmoAct2: Action Reasoning Models for Real-world Deployment}, 
      author={Haoquan Fang and Jiafei Duan and Donovan Clay and Sam Wang and Shuo Liu and Weikai Huang and Xiang Fan and Wei-Chuan Tsai and Shirui Chen and Yi Ru Wang and Shanli Xing and Jaemin Cho and Jae Sung Park and Ainaz Eftekhar and Peter Sushko and Karen Farley and Angad Wadhwa and Cole Harrison and Winson Han and Ying-Chun Lee and Eli VanderBilt and Rose Hendrix and Suveen Ellawela and Lucas Ngoo and Joyce Chai and Zhongzheng Ren and Ali Farhadi and Dieter Fox and Ranjay Krishna},
      year={2026},
      eprint={2605.02881},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2605.02881}, 
}