MolmoAct2: Action Reasoning Models for Real-world Deployment

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

• The paper introduces an open framework MolmoAct2 that integrates specialized embodied reasoning and continuous control, achieving state-of-the-art performance in real-world robotics deployment.
• The paper employs a novel two-stage action tokenization and adaptive depth reasoning (MolmoAct2-Think) to reduce inference latency, with ablation studies demonstrating significant performance improvements on diverse tasks.
• The paper verifies through practical evaluations that MolmoAct2 delivers robust zero-shot and fine-tuned performance, enhancing task success and trajectory quality across heterogeneous robot platforms.

MolmoAct2: Action Reasoning Models for Real-world Deployment

Motivation and System Overview

MolmoAct2 addresses critical barriers in scaling vision-language-action (VLA) models for practical, real-world robotics. Existing VLAs are hindered by closed models, lack of extensible data and code, high inference latency from explicit reasoning modules, and an inability to deploy out-of-the-box on affordable platforms with reliable success rates. MolmoAct2 is a fully open VLA framework that integrates specialized embodied reasoning, efficient continuous action prediction, adaptive spatial reasoning, and rapid fine-tuning, designed for reproducible, extensible, and robust deployment on heterogeneous robot platforms.

Figure 1: MolmoAct2 system overview: diverse data collection, large-scale model pretraining, adaptive-depth action reasoning, and multi-platform out-of-the-box deployment.

Data Collection and Embodiment Diversity

Central to MolmoAct2 is a large, curated set of robot manipulation datasets spanning various axes of cost, embodiment, and scene diversity:

• MolmoAct2-BimanualYAM: 720 hours of teleoperated, high-fidelity bimanual trajectory data, with real-world task, object, and scene diversity, sets a new scale for open bimanual robotics.

  Figure 2: Low-cost, standardized, and reproducible BimanualYAM data collection setup.

• MolmoAct2-DROID: A filtered subset from DROID, enforcing only high-quality, language-annotated, and contiguous action segments, supporting robust deployment on Franka platforms.
• MolmoAct2-SO100/101: Diverse community-contributed data from low-cost robots, filtered with multi-stage validity, quality, and reward-gated pipelines, enhancing generalization to out-of-distribution settings.
• Additional public academic datasets and a multimodal corpus further support backbone pretraining and model transferability.

  Figure 3: Data mixture aggregation: robot, academic, web-scale multimodal, and embodied reasoning datasets supporting robust and diverse training.

Model Architecture: VLA-Backbone with Flow-Matching Action Expert

Molmo2-ER Backbone

MolmoAct2 leverages Molmo2-ER, a vision-LLM tailored for embodied spatial reasoning. Molmo2-ER is trained with a specialize-then-rehearse schedule on 3.3M diverse samples, combining synthetic, simulator, web, and real robot data. It excels at spatial referencing, multi-view correspondence, video spatial QA, and compositional and egocentric reasoning. This backbone achieves a 63.8% average on 13 embodied-reasoning benchmarks, outperforming closed and open alternatives, including surpassing Gemini Robotics ER-1.5 and GPT-5 in average performance.

Action Tokenization and Continuous Control

Robotic control is realized by a two-stage procedure: (1) an OpenFAST Tokenizer maps continuous robot trajectories into a discrete, compact token sequence, enabling seamless integration with the VLM training objective across heterogeneous embodiments; (2) a flow-matching continuous action expert is grafted onto the VLM via per-layer key-value cache connections. Unlike prior work conditioning only on final hidden states, MolmoAct2’s expert cross-attends to the entire hierarchy of VLM features, improving geometric and semantic grounding.

Figure 4: MolmoAct2 pipeline: per-layer KV connections enable the action expert to condition on the complete VLM context; the expert is trained with flow matching to denoise and generate continuous robot action trajectories.

In training, discrete and continuous action outputs are co-supervised. The use of knowledge insulation ensures VLM gradients remain focused, isolating the flow loss to the expert and adapter projections. Systematic ablations show that per-layer KV connections and multiple flow samples ($K=8$) result in non-trivial average gains on downstream manipulation tasks.

Adaptive Depth Reasoning: MolmoAct2-Think

MolmoAct2-Think extends MolmoAct2 with adaptive discrete depth reasoning, significantly reducing inference token cost while maintaining geometric scene grounding. It predicts a discretized $10\times10$ grid of depth codes per frame, but, leveraging temporal staticness, generatively updates only regions evidencing visual change.

Figure 5: MolmoAct2-Think enhances action generation by autoregressively updating only changed spatial regions of the depth-token grid, keeping geometrical context current at minimal compute overhead.

This mechanism allows MolmoAct2-Think to significantly reduce rollout latency compared to full-frame reasoning methods while conferring interpretability and improved success rates, especially on long-horizon and generalization-sensitive tasks.

Empirical Results and Numerical Analysis

MolmoAct2 achieves strong empirical performance across benchmarks representing the spectrum of simulation and real-world deployment constraints.

• Molmo2-ER: Achieves state-of-the-art open-weight embodied reasoning performance on 9/13 industry benchmarks (overall 63.8%, +17pp over starting backbone).
• Out-of-the-box Zero-shot: On DROID and SO-100/101 platforms, MolmoAct2 achieves superior zero-shot task success (87.1% on real-world DROID tasks, +38.7pp compared to next best), with large margins even on low-cost robots.
• Fine-tuning Adaptation: On LIBERO and RoboEval, MolmoAct2 fine-tunes to 97.2% and 44.3% average success, exceeding strong fine-tuned baselines (e.g., $\pi_{0.5}$-DROID) by up to 3.8–15%.
• MolmoAct2-Think: Adaptive depth reasoning yields a further +0.9pp averaged improvement (98.1% vs 97.2%) on challenging manipulation suites.
• Trajectory Quality: MolmoAct2 outperforms baselines on non-task success metrics (completion time, trajectory/joint path length, jerk), essential for real-world safety and reliability.

  Figure 6: RoboEval success rates — MolmoAct2 dominates strong control and VLA baselines on task success and trajectory quality metrics.

  

  Figure 7: Real-world task adaptation: with few demonstrations, MolmoAct2 fine-tunes to outperform four baselines by 15% on a diverse suite of natural settings and mobile tasks.

• Inference Efficiency: With caching and CUDA Graph optimizations, MolmoAct2 achieves a 2.42x speedup, reaching 55.8Hz (Figure 8), demonstrating practical control rates for real-time deployment even with onboard hardware.

  Figure 8: Optimized inference throughput after cache and CUDA Graph integration for MolmoAct2 and MolmoAct2-Think.

Practical and Theoretical Implications

MolmoAct2 demonstrates that system-level openness—models, training code, data, and tokenizers—can be reconciled with performance, reproducibility, and efficient real-world deployment of generalist robotic policies. The architecture establishes best practices for action reasoning: dense vision-language and continuous control coupling, temporal-adaptive latent state reasoning, and resource-constrained fine-tuning. The strong performance on out-of-distribution real-world tasks with affordable hardware substantially lowers barriers for reproducible robotic research and application.

The release of the largest open bimanual dataset, curated multimodal mixtures, and a transparent action token representation enables comparative and extensible evaluation, critical for community progress on generalist manipulation.

Future Perspectives

MolmoAct2 provides several avenues for future research:

• Extending flow-matching continuous action experts to support multi-agent and long-horizon planning in compositional, dynamic environments.
• Deeper integration of temporally and spatially adaptive intermediate reasoning, potentially leveraging attention-efficient or speculative decoding approaches for further acceleration.
• Systematic investigation of knowledge insulation and cross-modality interaction for improved generalization and robustness.
• Open-ended transfer to highly constrained, embedded robotic platforms and further democratization of real-world AI manipulation.

Conclusion

MolmoAct2 establishes a robust methodology and resource set for open, scalable, and high-performance action reasoning in robotics. The results conclusively show that with appropriate backbone specialization, data curation, architecture design, and reasoning optimization, fully open models can surpass proprietary and closed alternatives, even on real-world deployment metrics. The framework sets a new standard for reproducible, actionable VLA research, supporting future advances in embodied AI (2605.02881).