Paper 2605.02881
MolmoAct2: Action Reasoning Models for Real-world Deployment
- Published
- May 2026
- Research lab
- Independent
- Citations
- 23
- GitHub
- 697 stars
01 In brief
Summary
MolmoAct2 is a fully open Vision-Language-Action (VLA) model family for real-world robot deployment, built on the Molmo2-ER backbone, a VLM specialized for spatial and embodied reasoning trained on a 3.3M-sample corpus.
It introduces three new robot datasets: MolmoAct2-BimanualYAM (720 hours, largest open bimanual dataset), MolmoAct2-DROID (quality-filtered Franka subset), and MolmoAct2-SO100/101 (quality-filtered community data).
The MolmoAct2-FAST Tokenizer, an open-weight action tokenizer, compresses continuous actions into discrete tokens.
The architecture uses a DiT-style flow-matching action expert conditioned via per-layer KV conditioning from the VLM.
MolmoAct2-Think, an adaptive-depth reasoning variant, re-predicts depth tokens only for changed scene regions, reducing latency.
In extensive evaluations across 7 benchmarks, MolmoAct2 outperforms baselines like π0.5, and Molmo2-ER surpasses GPT-5 and Gemini Robotics ER-1.5 on 13 embodied-reasoning benchmarks.
All weights, code, and data are released.
Key results include 87.1% success on real-world DROID tasks, 97.2% on LIBERO, and 50.7% robustness under perturbations, with inference speeds up to 55.79 Hz after optimizations.
The model supports out-of-the-box deployment on YAM, SO-100/101, and DROID Franka, and fine-tunes efficiently for new tasks and embodiments.
Ablations confirm the benefits of the embodied-reasoning backbone, per-layer KV conditioning, multiple flow samples, and adaptive depth reasoning.
Limitations include open-loop action chunks and zero-shot deployment limited to trained embodiments.
The work aims to democratize VLA research by providing complete open resources for reproducible and deployable robot learning.
- MolmoAct2 is built on Molmo2-ER, a VLM specialized for spatial and embodied reasoning, trained on a 3.3M-sample corpus.
- Releases three new robot datasets: MolmoAct2-BimanualYAM (720 hours, largest open…
02 From the paper
Abstract
Vision-Language-Action (VLA) models aim to provide a single generalist controller for robots, but today's systems fall short on the criteria that matter for real-world deployment. Frontier models are closed, open-weight alternatives are tied to expensive hardware, reasoning-augmented policies pay prohibitive latency for their grounding, and fine-tuned success rates remain below the threshold for dependable use. We present MolmoAct2, a fully open action reasoning model built for practical deployment, advancing its predecessor along five axes. We introduce MolmoER, a VLM backbone specialized for spatial and embodied reasoning, trained on a 3.3M-sample corpus with a specialize-then-rehearse recipe. We release three new datasets spanning low-to-medium cost platforms, including MolmoAct2-BimanualYAM, 720 hours of teleoperated bimanual trajectories that constitute the largest open bimanual dataset to date, together with quality-filtered Franka (DROID) and SO100/101 subsets. We provide OpenFAST, an open-weight, open-data action tokenizer trained on millions of trajectories across five embodiments. We redesign the architecture to graft a flow-matching continuous-action expert onto a discrete-token VLM via per-layer KV-cache conditioning. Finally, we propose MolmoThink, an adaptive-depth reasoning variant that re-predicts depth tokens only for scene regions that change between timesteps, retaining geometric grounding at a fraction of prior latency. In the most extensive empirical study of any open VLA to date, spanning 7 simulation and real-world benchmarks, MolmoAct2 outperforms strong baselines including Pi-05, while MolmoER surpasses GPT-5 and Gemini Robotics ER-1.5 across 13 embodied-reasoning benchmarks. We release model weights, training code, and complete training data. Project page: https://allenai.org/blog/molmoact2