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

Figures (8)

• Figure 1: Overview of MolmoAct2. MolmoAct2 is a fully open action reasoning model for real-world deployment. From a suite of high-quality robot datasets that we collect, filter, and curate at scale across three platforms spanning the low-to-medium cost range (left), we train MolmoAct2 and its adaptive-depth reasoning variant MolmoAct2-Think, coupled to the VLM backbone through a novel action expert connection (center). The resulting models deploy out-of-the-box on bimanual YAM, SO-100/101, and DROID Franka, and adapt to in-the-wild tasks such as cleaning up, washing dishes, wetlab automation, and pouring tea (right).
• Figure 2: Overview of the MolmoAct2 training data. Our data mixture combines public academic robot datasets, MolmoAct2-BimanualYAM Dataset, MolmoAct2-DROID Dataset, MolmoAct2-SO100/101 Dataset, multimodal web data, and embodied-reasoning data.
• Figure 3: MolmoAct2-BimanualYAM Dataset collection setup. Our standardized MolmoAct2-BimanualYAM Dataset collection setup. Every component is readily available for purchase, and the total cost of the entire setup is under $6,000 USD.
• Figure 4: Overview of MolmoAct2. Image observations, language instructions, and robot states are tokenized and processed by a pre-trained VLA backbone with self-attention layers. Post-training attaches a DiT-style action expert with the same number of layers as the backbone. At each layer, the backbone key and value tensors are projected and reused as the key and value inputs to the action expert's cross-attention, forming a per-layer KV connection from visual-language context to continuous control. The expert is trained with flow matching to denoise a noisy action trajectory into a continuous robot trajectory. During training, the backbone is also supervised with next-token prediction over discrete action tokens, while the target action-token span is masked from the expert so continuous action prediction cannot condition on the ground-truth discrete actions.
• Figure 5: Overview of MolmoAct2-Think. MolmoAct2-Think augments the MolmoAct2 action-generation pipeline with adaptive depth-token reasoning, reusing cached depth codes for static regions and regenerating depth codes for changed regions before conditioning the action expert.