FlowVLA: Visual Chain of Thought-based Motion Reasoning for Vision-Language-Action Models

Many Vision-Language-Action (VLA) models are built upon an internal world model trained via next-frame prediction $v_t \rightarrow v_{t+1}$''. However, this paradigm attempts to predict the future frame's appearance directly, without explicitly reasoning about the underlying dynamics. \textbf{This lack of an explicit motion reasoning step} often leads to physically implausible visual forecasts and inefficient policy learning. To address this limitation, we introduce the \textbf{Visual Chain of Thought (Visual CoT)}, a paradigm that compels the model to first reason about \textbf{motion dynamics} before generating the future frame. We instantiate this paradigm by proposing \textbf{FlowVLA}, an autoregressive Transformer that explicitly materializes this reasoning process as $v_t\to f_t\to v_{t+1}$‘’, where $f_t$ is an intermediate optical flow prediction that inherently encodes motion. By forcing the model to first follow the motion plan encoded by $f_t$, this process inherently $\mathbf{\text{aligns}}\mathbf{\text{the}}\mathbf{\text{pre-training}}\mathbf{\text{objective}}\mathbf{\text{of}}\mathbf{\text{dynamics}}\mathbf{\text{prediction}}\mathbf{\text{with}}\mathbf{\text{the}}\mathbf{\text{downstream}}\mathbf{\text{task}}\mathbf{\text{of}}\mathbf{\text{action}}\mathbf{\text{generation.}}$ We conduct experiments on challenging robotics manipulation benchmarks, as well as real-robot evaluations. Our FlowVLA not only generates $\mathbf{\text{more}}\mathbf{\text{coherent}}\mathbf{\text{and}}\mathbf{\text{physically}}\mathbf{\text{plausible}}\mathbf{\text{visual}}\mathbf{\text{predictions}}$, but also achieves state-of-the-art policy performance with $\mathbf{\text{substantially}}\mathbf{\text{improved}}\mathbf{\text{sample}}\mathbf{\text{efficiency}}$, pointing toward a more principled foundation for world modeling in VLAs. Project page: https://irpn-lab.github.io/FlowVLA/