FoMoVLA: Bridging Visual Foresight and Motion Constraints for Vision-Language-Action Models

Abstract

Vision-Language-Action (VLA) models have achieved impressive results in visuomotor policy learning, yet remain fundamentally reactive, mapping current observations and language to actions without explicit forward prediction of world dynamics. Although existing approaches incorporate future visual prediction, they still face considerable challenges from visual redundancy and the absence of fine-grained motion cues for action. We posit that future scene anticipation and geometric motion tracking are complementary yet inseparable: the former specifies the goal visual state, while the latter provides the continuous motion dynamics toward it. From this perspective, we propose FoMoVLA, a framework that augments VLA representations with explicit spatio-temporal supervision by jointly learning future visual foresight and 2D motion tracking. FoMoVLA introduces compact foresight tokens to encode future visual states, decodes temporal 2D point trajectories to model geometric motion, and couples both through a lightweight future-conditioned cross-attention module that enables mutually reinforced reasoning between anticipated states and point dynamics. Extensive experiments on LIBERO, RoboCasa GR-1 Tabletop, and LIBERO-Plus demonstrate state-of-the-art performance and strong zero-shot generalization.

Method

Overview of FoMoVLA. Given a current observation and language instruction, the VLM backbone processes image tokens (with text tokens placed first for goal-aware conditioning) alongside K learnable <Foresight> tokens.

Results

LIBERO Benchmark

LIBERO evaluation results. The full model achieves the highest average success rate across all 4 task suites.

Zero-Shot OOD Robustness (LIBERO-Plus)

Out-of-distribution robustness on LIBERO-Plus (%). All methods evaluated zero-shot without fine-tuning.

RoboCasa GR-1 Tabletop

Average success rate on RoboCasa GR-1 Tabletop (24 tasks, 50 rollouts/task).

Visualization

Qualitative Results on LIBERO

Qualitative results of FoMoVLA. FoMoVLA effectively learns future features and point tracking.

Effect of FCCA

Effect of FCCA on point tracking. Conditioning point tracking on visual foresight yields trajectories aligned with action dynamics.

Demo Videos

LIBERO & LIBERO-Plus

RoboCasa GR-1 Tabletop

Citation

@misc{li2026fomovlabridgingvisualforesight,
      title={FoMoVLA: Bridging Visual Foresight and Motion Guidance for Vision-Language-Action Models},
      author={Wei Li and Peijin Jia and Yuan Ma and Xuefeng Jiang and Titong Jiang and Sheng Sun and Yujian Li and Xin Wen and Han Hong and Zhikang Liu and Bailin Li and Kun Zhan},
      year={2026},
      eprint={2607.14739},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2607.14739}
}