MolmoAct2: An Open Visual-Language-Action Reasoning Model for Real-World Deployment

MolmoAct2 is a fully open-source Visual-Language-Action (VLA) model developed by the Allen AI team, designed specifically for real-world deployment. Through five core innovations (MolmoER backbone network, three new datasets, OpenFAST action tokenizer, flow-matching continuous action expert architecture, and MolmoThink adaptive reasoning), it outperforms strong baselines like Pi-05 on 7 simulation and real-world benchmarks, providing an open and scalable research platform for the robotics field.

Current VLA systems face four major challenges:

• Closedness: Most cutting-edge models are closed-source, making customization and optimization impossible
• Hardware dependency: Open-source solutions are tied to expensive specialized hardware
• Latency issues: Enhancing reasoning capabilities sacrifices real-time performance
• Success rate bottleneck: Even after fine-tuning, it's still difficult to meet the threshold for reliable deployment These pain points restrict the implementation of VLA technology in industrial and service robotics fields.

1. MolmoER Backbone: Optimized for spatial and embodied reasoning, using the "Specialize-then-Rehearse" training strategy, outperforming closed-source models like GPT-5
2. Three New Datasets: Covering bimanual operations (MolmoAct2-BimanualYAM), high-signal-to-noise Franka subsets, and low-cost platform SO100/101 subsets
3. OpenFAST Tokenizer: An open-source action discretization tool that breaks the limitations of closed-source/platform binding
4. Flow-Matching Architecture: Integrates discrete token VLM with continuous action experts to achieve fine-grained control
5. MolmoThink Reasoning: Adaptively updates areas of scene changes to reduce latency

• Evaluation covers 7 simulation/real-world benchmarks, outperforming the strong baseline Pi-05
• MolmoER outperforms GPT-5 and Gemini Robotics ER-1.5 on 13 embodied reasoning benchmarks
• Cross-platform generalization: Adapts from Franka to low-cost SO100/101
• Fully open-source: Model weights, training code, and data are all publicly available

• Training Strategy: Specialize-then-Rehearse (first specialize in training robot tasks, then rehearse general data to avoid overfitting)
• Flow-Matching Integration: Seamlessly connects continuous action experts with discrete token VLM through KV cache conditioning
• Adaptive Reasoning: Scene change detection + sparse updates, maintaining accuracy while reducing latency

• Lower research barriers: No need to rely on closed-source APIs or specialized hardware
• Promote standardization: Tools like OpenFAST establish community reuse standards
• Accelerate industrial implementation: Adapt to multi-cost platforms, paving the way for industrialization

Limitations:

• The scale of bimanual datasets is still smaller than industrial-grade
• Simulation-to-real transfer is not fully resolved
• Performance on long-range complex tasks needs verification

Future Directions:

• Expand data scale and diversity
• Strengthen simulation-to-real transfer technology
• Improve long-range task planning capabilities

MolmoAct2 is an important milestone in the development of open-source VLA models. Not only does it outperform baselines in performance, but it also provides a fully open research platform. Its open-source nature will promote progress in the fields of robot learning and embodied intelligence, providing key infrastructure for industry implementation.