Official Review · ICLR 2026

This paper presents XR-1, a unified vision-language-action framework that introduces the Unified Vision-Motion Codes (UVMC) to jointly quantize and align visual dynamics with motor actions. The authors conduct extensive real-world experiments with over 12,000 rollouts across six different robot embodiments, demonstrating strong cross-task and cross-embodiment generalization. - The paper presents extensive real-world experiments, validating XR-1 with over 12,000 rollouts across six different robot embodiments. - This paper presents XR-1, a unified vision-language-action framework that introduces the Unified Vision-Motion Codes (UVMC) to jointly quantize and align visual dynamics with motor actions. The authors conduct extensive real-world experiments with over 12,000 rollouts across six different robot embodiments, demonstrating strong cross-task and cross-embodiment generalization.The proposed method is novel, introducing the Unified Vision-Motion Codes (UVMC) that jointly quantize and encode both actions and observations. - The paper does not appear to include simulation experiments. Although the real-world evaluation is rich and convincing, an evaluation on a standard public benchmark would provide a more comprehensive assessment of performance and facilitate reproducibility for the research community. *I strongly suggest author take this into consideration*. - Analytical justification of UVMC: The core innovation UVMC lacks deeper justification beyond ablation studies. The observed improvements could potentially be influenced by hyperparameter choices or implementation factors. I recommend adding more analytical or theoretical explanations to strengthen the evidence for UVMC’s effectiveness. - Interpretability of UVMC: The interpretability of the UVMC alignment mechanism could also be improved. Although the KL alignment encourages visual and motion features to share the same codebook, the paper does not analyze how these codes are organized or clustered in semantic space. For example, which codes may correspond to grasping, rotation, or translation. As a result, the semantic consistency of UVMC remains an assumption rather than an explicitly demonstrated property. - I am somewhat unclear about the UVMC design. During reconstruction, both observation and action encoders take additional inputs such as c_t and l, o. Could the authors provide more explanation on the rationale behind this design? - The latent variables are defined as $z_{\text{vis}} = E_{\text{vis}}(c_t, c_{t+h})$ and $z_{\text{mo}} = E_{\text{mo}}(a_{t:t+h}, m_{t:t+h}).$ This paper propose to let the latent representation be as close as possible. However, therotically, the same $(a_{t:t+h}, m_{t:t+h})$ can correspond to very different $(c_t, c_{t+h})$, since the action sequence and motor states only describe the robot itself, while the visual inputs capture the entire scene, including background and objects, which can be changed arbitrarily across episodes. Directly aligning th