MolmoAct2: Action Reasoning Models for Real-World Deployment

Authors: Haoquan Fang, Jiafei Duan, Donovan Clay, Sam Wang, Shuo Liu, Weikai Huang, Xiang Fan, Wei-Chuan Tsai, Shirui Chen, Yi Ru Wang, Shanli Xing, Jaemin Cho, Jae Sung Park, Ainaz Eftekhar, Peter Sushko, Karen Farley, Angad Wadhwa, Cole Harrison, Winson Han, Ying-Chun Lee, Eli VanderBilt, Rose Hendrix, Suveen Ellawela, Lucas Ngoo, Joyce Chai, Zhongzheng Ren, Ali Farhadi, Dieter Fox, Ranjay Krishna · CoRL 2026 2026

Why it matters

Most capable robot policies are closed, and the open ones usually release weights without the data and recipe behind them, or assume expensive robots. MolmoAct2 publishes all three: model weights, training code and the full training mixture, including about 720 hours of new bimanual teleoperation data on YAM arms whose whole setup the authors put at under 6,000 US dollars. The paper is the reference for how that release was built and how well it works, and CoRL 2026 chose it as a spotlight. For a reader of this catalogue the practical part is that two of its fine-tuned checkpoints run without further training on hardware many labs and hobbyists already own: a DROID-style Franka arm and the SO-100 or SO-101 arm. The paper also shows ideas other model builders can reuse: an openly trained FAST-style action tokenizer, a filtering pipeline that turns 1,222 crowd-sourced LeRobot datasets into usable training data, and an action expert conditioned on the backbone's keys and values at every layer, which beat conditioning on hidden states in the authors' ablation. The model itself, with its checkpoints and deployment guides, has its own page in our catalogue; this page covers the paper's method and results.

Problem

Vision-language-action models promise one controller for many robots, but the paper names four obstacles to real deployment. The strongest models are closed, and the few open-weight ones withhold their data and training recipe, so nobody can reproduce or extend them. The open models that run out of the box are tied to expensive or specialised robots. Policies that reason before acting, for example by predicting depth, goal images or trajectories, spend hundreds of tokens per step and become too slow for closed-loop control. And even after fine-tuning, success on realistic tasks stays below what dependable use needs. Public robot data does not fix this on its own: Open X-Embodiment varies in quality, much of DROID is idle or repetitive, community SO-100 uploads carry placeholder labels such as lerobot_test, and no open dataset offered high-quality bimanual data on the platform the authors wanted to deploy on.

Proposed solution

The authors first fine-tune their Molmo2 vision-language model into Molmo2-ER on a corpus of about 3.3 million embodied-reasoning examples (pointing, spatial and video questions, multi-view correspondence). They collect a bimanual YAM dataset of 34,500 demonstrations over more than 28 tasks, and filter community SO-100/101 uploads and DROID by structure, licence and quality, re-labelling the instructions with an open VLM. Training then runs in three stages. Pre-training teaches the backbone to emit discrete action tokens from MolmoAct2-FAST Tokenizer, an open re-implementation of FAST with a 2,048-token vocabulary trained on one million action sequences from five embodiments. Post-training attaches a DiT-style action expert with the same 36 layers as the backbone; each expert layer cross-attends to the keys and values of the matching backbone layer and is trained with flow matching, while the backbone keeps its token loss. Fine-tuning then adapts the model to one robot, producing checkpoints for bimanual YAM, DROID Franka, SO-100/101 and LIBERO. MolmoAct2-Think adds a 10 by 10 grid of depth tokens before acting and re-predicts only the cells whose image patches changed since the last step. Caching and CUDA Graphs speed up inference.

Key insights

A backbone trained for embodied reasoning helps action learning directly: swapping Molmo2 for Molmo2-ER raises LIBERO-Long success from 77.6% to 83.6% with the same discrete-action setup
The action expert works better when each of its layers reads the matching backbone layer's keys and values: 95.9% on LIBERO against 94.0% when it conditions on hidden states
Reasoning does not have to cost a full re-prediction every step: MolmoAct2-Think re-predicts depth tokens only for image cells that changed, keeping the geometric grounding at a fraction of the latency
Curation matters as much as scale for community data: structural checks, licence checks and a quality gate turn 1,222 crowd-sourced SO-100/101 datasets into training data, and re-labelling with an open VLM doubles the number of unique instructions
The continuous path is the fast one: decoding the same action chunk as discrete tokens runs at 14.17 Hz against 55.79 Hz for the flow-matching expert, so the expert is the default for deployment

Main results

Out of the box on a real DROID Franka, with randomly placed cameras and unseen objects and scenes, MolmoAct2-DROID averages 87.1% over five pick-and-place tasks (15 trials each), against 48.4% for MolmoBot and 45.2% for π0.5-DROID
Out of the box on a real SO-100 arm, MolmoAct2-SO100/101 scores 56.7% over five tasks with partial credit, against 45.3% for a π0 the authors fine-tuned on the same data and 2.3% for SmolVLA
In simulation without fine-tuning, MolmoAct2-DROID averages 37.7% on MolmoSpaces against 34.5% for π0.5-DROID, and 20.6% on MolmoBot against 10.0%
After fine-tuning it reaches 97.2% on LIBERO (98.1% for MolmoAct2-Think), against 96.9% for π0.5 and 86.6% for the earlier MolmoAct, and 44.3% on the bimanual RoboEval benchmark against 40.5% for π0.5
Fine-tuned on 8 real bimanual YAM tasks with 50 trials each, it averages 50.1% and wins 7 of the 8, 15 points above the runner-up, OpenVLA-OFT
Molmo2-ER averages 63.8% across 13 embodied-reasoning benchmarks, ahead of Gemini Robotics-ER 1.5 Thinking (61.3%) and GPT-5 (57.9%), and 17 points above the Molmo2 model it starts from
On one H100, caching and CUDA Graphs raise the control rate from 23.02 Hz to 55.79 Hz for MolmoAct2 and from 8.04 Hz to 12.71 Hz for MolmoAct2-Think

Limitations

The licence is clear for the code and less so for the weights: the repository's LICENSE file is Apache-2.0, and the README says the model is licensed under Apache 2.0 and intended for research and educational use under Ai2's Responsible Use Guidelines, but the Hugging Face cards of the MolmoAct2 checkpoints state no licence (only the Molmo2-ER card does). The robot dataset cards say Apache-2.0, while some re-hosted Molmo2-ER training sets keep their sources' terms, among them VST-P under the non-commercial CC BY-NC 4.0
The policy executes fixed action chunks open-loop: nothing enforces smooth motion across chunk boundaries, it cannot react to contact or disturbance within a chunk, and the authors note that the 55.79 Hz figure is amortised chunk throughput, not closed-loop reactivity
Zero-shot use is limited to the three platforms it was trained for (bimanual YAM, SO-100/101 and DROID Franka); other robots, hands, mobile manipulators or humanoids need fine-tuning on their own demonstrations, and the README says out-of-the-box performance has only been verified on SO-100 and Franka DROID, for simple tasks within the training distribution
Results are the authors' own: the real zero-shot tests use 15 trials per task, the SO-100 scores give partial credit, two simulation benchmarks (MolmoSpaces and MolmoBot) and the MolmoBot baseline come from Ai2, the π0 SO-100 baseline was fine-tuned by the authors, and the robustness table labels its row MolmoAct2-Think while the text attributes the result to MolmoAct2
Training is far beyond a single lab GPU: pre-training alone used about 5,760 H100 GPU hours and post-training about 2,304, the checkpoints are about 22 GB each, and the README notes that bfloat16 inference needs patches to the model code

Field impact

MolmoAct2 is one of the few vision-language-action models whose weights, code and complete training data are all public, so other groups can retrain, audit or extend every stage rather than only fine-tune a released checkpoint. It moves the open baseline onto affordable hardware: the authors report out-of-the-box results on an SO-100 arm and a DROID Franka in unseen scenes, and fine-tuned results on a bimanual YAM setup outside the lab. Two design results are likely to travel beyond this model: conditioning the action expert on the backbone's per-layer keys and values, and making reasoning tokens adaptive so their cost follows how much the scene changes. The released bimanual YAM data, described by the authors as the largest open bimanual dataset to date, and the filtered SO-100/101 and DROID subsets are resources in their own right. Many of the comparisons use the authors' own benchmarks and baselines, so independent replication will decide how far the margins hold.

Abstract

Vision-Language-Action (VLA) models aim to provide a single generalist controller for robots, but today's systems fall short 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, MolmoAct along five axes. (1) MolmoAct2 is built on top of our new Molmo2-ER, a VLM backbone specialized for spatial and embodied reasoning, trained on a 3.3M-sample corpus with a specialize-then-rehearse recipe. (2) We release three new robot datasets spanning low-to-medium cost platforms: MolmoAct2-BimanualYAM Dataset, 720 hours of teleoperated bimanual trajectories that constitute the largest open bimanual dataset to date; MolmoAct2-DROID Dataset, a quality-filtered Franka subset of DROID; and MolmoAct2-SO100/101 Dataset, a quality-filtered SO-100/101 subset. (3) We train and release MolmoAct2-FAST Tokenizer, an open-weight, open-data action tokenizer trained on millions of trajectories across five embodiments. (4) We design a new VLA architecture to graft the discrete-token VLM into the flow-matching continuous-action expert via per-layer key-value (KV) conditioning. (5) we propose MolmoAct2-Think, 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 π0.5, while Molmo2-ER surpasses GPT-5 and Gemini Robotics ER-1.5 across 13 embodied-reasoning benchmarks. We release model weights, training code, and complete training data.

Open Source Robotics