This assessment is based on version 1 of this paper. Version 2 is now available on arXiv — the authors may have revised their methods, results, or conclusions.
MolmoAct2: Action Reasoning Models for Real-world Deployment
Haoquan Fang, Jiafei Duan, Donovan Clay, Sam Wang, Shuo Liu, Weikai Huang, Xiang Fan, Wei-Chuan Tsai
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
AI Impact Assessments
(1 model)Scientific Impact Assessment: MolmoAct2
1. Core Contribution
MolmoAct2 presents a fully open (weights, code, data) vision-language-action model system designed for practical robot deployment. The paper advances along five interconnected axes: (1) Molmo2-ER, a VLM backbone specialized for embodied reasoning via a 3.3M-sample corpus and a specialize-then-rehearse training recipe; (2) three new robot datasets, including the 720-hour bimanual YAM dataset (claimed largest open bimanual dataset); (3) OpenFAST, an open-weight/open-data action tokenizer; (4) a novel architecture grafting a flow-matching continuous-action expert onto a discrete VLM via per-layer KV-cache conditioning; and (5) MolmoAct2-Think, an adaptive-depth reasoning mechanism that selectively re-predicts depth tokens only for changed scene regions.
The paper addresses a genuine multi-faceted problem: frontier VLAs are closed-source, open alternatives require expensive hardware, reasoning-augmented policies are too slow, and fine-tuned success rates remain inadequate. The contribution is therefore a systems-level integration rather than a single algorithmic breakthrough.
2. Methodological Rigor
Strengths in experimental design: The evaluation is remarkably comprehensive—7 simulation and real-world benchmarks for the VLA, 13 embodied-reasoning benchmarks for Molmo2-ER, spanning multiple embodiments (bimanual YAM, Franka/DROID, SO-100/101). The paper includes confidence intervals for simulation results, systematic ablations on LIBERO isolating each design choice (backbone, KV-connection type, flow sample count, fine-tuning recipe, depth-aware training), and inference latency measurements. The ablation structure is well-designed, with each table changing one variable at a time.
Concerns: Several aspects warrant scrutiny. The real-world evaluations use only 15 trials per task for out-of-the-box deployment and 50 trials for fine-tuning experiments—while standard, these sample sizes limit statistical power for detecting meaningful differences. The LIBERO benchmark appears near saturation (97-98% for multiple methods), making marginal improvements harder to interpret. Some comparisons may not be perfectly apples-to-apples: π0.5-DROID is compared against MolmoAct2-DROID, but differences in training data composition, compute budget, and model size could confound conclusions. The paper does not provide detailed failure analysis or systematic characterization of failure modes, which would be important for the "real-world deployment" framing.
The Molmo2-ER evaluation against GPT-5 and Gemini Robotics ER-1.5 is notable but should be interpreted carefully—the benchmarks may favor the specific spatial/embodied skills the model was trained on, and the "overall average" across 13 benchmarks gives equal weight to each regardless of downstream importance.
3. Potential Impact
Immediate practical impact: The fully open release (weights, code, training data) is perhaps the most impactful aspect. The robotics community has been hampered by closed-source frontier models, and MolmoAct2 provides a reproducible, extensible baseline that supports three cost-accessible platforms. The BimanualYAM dataset (720 hours) fills an important gap in bimanual manipulation data.
Technical contributions with lasting value: The per-layer KV-cache conditioning mechanism for connecting VLM backbones to continuous-action experts is architecturally elegant and well-ablated. The adaptive-depth reasoning in MolmoAct2-Think, while showing modest quantitative gains, introduces a principled approach to reducing reasoning latency that exploits temporal coherence in robot observation streams. OpenFAST provides a reproducible action tokenizer with transparent training data—a practical infrastructure contribution.
Broader influence: This work could accelerate VLA research in labs without access to expensive hardware or proprietary models. The specialize-then-rehearse recipe for Molmo2-ER demonstrates an efficient paradigm for injecting domain-specific capabilities into foundation models without full retraining.
4. Timeliness & Relevance
The paper directly addresses the current bottleneck in robotics foundation models: the gap between impressive closed-source demonstrations and what open-source researchers can reproduce and deploy. The timing is excellent—the field is actively seeking open alternatives to π0.5 and Gemini Robotics. The focus on low-to-medium cost platforms (SO-100/101 at ~6,000) responds to a real accessibility need.
The adaptive reasoning approach in MolmoAct2-Think addresses the emerging tension between reasoning-augmented policies and real-time control requirements—a bottleneck that will only intensify as reasoning becomes standard in robot policies.
5. Strengths & Limitations
Key Strengths:
Notable Limitations:
Additional Observations
The paper represents a significant engineering and systems contribution. The integration of discrete pre-training with continuous post-training via KV-cache conditioning is a clean architectural insight. The community data filtering pipeline for SO-100/101 (using TOPReward) is a practical contribution for the growing ecosystem of community-collected robot data.
Generated May 5, 2026
Comparison History (69)
MolmoAct2 has broader scientific impact across multiple dimensions. It addresses a fundamental challenge in robotics (generalist vision-language-action models) with contributions spanning architecture design, large-scale dataset release (720h bimanual data), open-weight models, and novel techniques (adaptive-depth reasoning, flow-matching action expert). Its fully open release of weights, code, and data maximizes community impact. The extensive evaluation across 7 benchmarks and 13 embodied-reasoning benchmarks, outperforming frontier models like GPT-5 and Pi-05, demonstrates exceptional breadth. ParkourFormer, while strong in its niche of humanoid parkour, addresses a narrower problem with more incremental architectural contributions.
MolmoAct2 has higher potential impact due to its broader scope and practical deployment focus. It advances along five axes simultaneously: a specialized VLM backbone outperforming GPT-5/Gemini on embodied reasoning benchmarks, the largest open bimanual dataset (720 hours), an open action tokenizer across five embodiments, a novel flow-matching architecture, and adaptive-depth reasoning reducing latency. Its fully open release (weights, code, data) across diverse hardware platforms democratizes access. The breadth of evaluation (7 sim/real benchmarks, 13 reasoning benchmarks) and multi-embodiment generalization give it wider cross-field impact than ERVLA's more focused contribution on CoT integration strategies.
Paper 2 (MolmoAct2) likely has higher scientific impact due to broader scope and stronger real-world deployment orientation: it delivers an open, end-to-end VLA system with multiple novel components (specialized embodied-reasoning VLM, open action tokenizer, hybrid discrete/continuous action modeling, adaptive-depth reasoning) plus major open datasets (notably a large bimanual set) and full releases of weights/code/data. Its claimed evaluation breadth (many sim+real benchmarks and reasoning suites) suggests wide cross-field influence. Paper 1 is novel and rigorous but more narrowly focused on short-term memory and a smaller benchmark.
Paper 1 likely has higher impact due to broader scope and stronger enabling contributions: an open, deployable VLA system plus multiple large-scale open datasets (including the largest open bimanual set), an open action tokenizer, architectural innovations for continuous control with KV-cache conditioning, and a latency-focused reasoning method (MolmoThink). It reports extensive evaluations across many benchmarks and releases weights, code, and full training data—maximizing reproducibility and downstream adoption across robotics and embodied AI. Paper 2 is novel and timely for scaling dexterous data, but its impact may be narrower and less infrastructural.
MolmoAct2 offers broader practical impact through five concrete technical advances: a specialized VLM backbone, the largest open bimanual dataset (720hrs), an open action tokenizer across 5 embodiments, a novel flow-matching architecture, and adaptive-depth reasoning reducing latency. Its fully open release (weights, code, data) on low-to-medium cost platforms lowers barriers significantly. While Embodied-R1.5 achieves strong benchmark results with its unified EFM and closed-loop framework, MolmoAct2's emphasis on deployability, hardware accessibility, open infrastructure, and the most extensive empirical evaluation of any open VLA gives it higher real-world and community impact.
MolmoAct2 demonstrates broader scientific impact through its fully open ecosystem (weights, code, data), five distinct technical contributions spanning backbone training, architecture design, adaptive reasoning, and the largest open bimanual dataset (720hrs). It covers 7 sim/real benchmarks and 13 embodied-reasoning benchmarks, outperforming frontier models like GPT-5 and Pi-05. Its open resources lower barriers for the entire community. World Pilot is innovative but narrower in scope, focusing primarily on world-model priors for VLAs with evaluation on fewer benchmarks.
Paper 2 presents a comprehensive, fully open foundation model for robotic action reasoning. By releasing a specialized VLM backbone, massive new bimanual datasets, an open-weight action tokenizer, and an adaptive reasoning architecture, it provides critical infrastructure for the robotics community. Its broad scope, extensive benchmarking against frontier models, and open-source nature suggest a much wider, field-level impact compared to Paper 1's more focused, albeit innovative, approach to dexterous manipulation.
Paper 1 has higher impact potential due to broader, more foundational contributions: an open, deployable VLA system with new backbone (MolmoER), large-scale datasets (notably the largest open bimanual set), an open action tokenizer, architectural innovations for combining discrete VLMs with continuous control, and a latency-reducing reasoning method (MolmoThink). Its extensive evaluation across many benchmarks plus full release of weights/data/code increases reproducibility and adoption across robotics and ML. Paper 2 is novel and practical for dexterous imitation from video, but is narrower in scope and less clearly a community-enabling platform.
MolmoAct2 has higher potential impact due to its comprehensive, multi-axis contributions to the VLA field: a new VLM backbone, the largest open bimanual dataset (720 hours), an open action tokenizer, a novel architecture combining flow-matching with discrete VLMs, and adaptive-depth reasoning reducing latency. Its extensive evaluation across 7+13 benchmarks surpassing frontier models (GPT-5, Pi-05), plus full open release of weights, code, and data, positions it to become foundational infrastructure. While RoboDream's compositional world model for data synthesis is innovative, MolmoAct2's breadth of contributions and open ecosystem approach will likely drive broader adoption and downstream research.
Paper 1 has higher potential impact due to its massive open-source contributions (models, enormous datasets, tokenizers, and code) in the highly relevant field of Vision-Language-Action models. It addresses critical real-world deployment bottlenecks like latency and hardware costs through novel architectures, such as adaptive-depth reasoning and flow-matching experts. Outperforming frontier models while fully open-sourcing the artifacts guarantees widespread community adoption and citations. While Paper 2 offers a rigorous theoretical framework, Paper 1 provides foundational infrastructure that will broadly accelerate embodied AI research.