Temporal Consistency-Constrained Vision–Action Latent Space Alignment and Policy Optimization

Real-world visuomotor control requires policies to reason over uncertain visual observations, action continuity, and long-horizon error accumulation. Existing vision-action policies often improve action generation but insufficiently align visual state evolution with action-induced latent dynamics. This paper proposes TCVA-PO, a temporal consistencyconstrained framework for vision-action latent space alignment and policy optimization. Its defining contribution is the two-stage coupling of cross-modal latent increments from paired demonstrations with a policy-induced next-latent transition penalty; pointwise alignment, reconstruction, critic regression, advantage weighting, and smoothness supply complementary representation and optimization structure. Experiments on six public robot manipulation datasets show that TCVA-PO consistently improves task performance, latent consistency, action smoothness, and long-horizon robustness over representative imitation and generative visuomotor baselines under heterogeneous visual-control conditions and noisy demonstration settings. These results demonstrate that temporally aligned latent dynamics provide a more stable basis for robust policy learning in complex uncertain robotic systems.