Real-Time EXPO-FT: Reinforcement Learning for Real-Time Vision-Language-Action Policies

Real robots do not pause for inference. The scale that makes a pretrained vision-language-action (VLA) model a strong behavior prior also inflates its inference latency — and the physical world keeps moving while the robot computes. The observation an action was inferred from is rarely the observation at the moment it executes, a distribution shift that erodes reliability exactly where it matters most. Methods like real-time chunking keep control smooth under this latency, but give the policy no mechanism to move beyond its training distribution. Real-Time EXPO-FT closes that gap: a reinforcement learning framework for fine-tuning VLA policies that are real-time by construction, so the reliability gains of RL and the reactivity of real-time execution come together rather than at each other’s expense.

• Slow generation, fast editing. Building on EXPO-FT, we decouple the two. A large VLA base policy proposes action chunks and is free to take as long as it needs; a small edit policy handles the fast, observation-conditioned decisions.
• Edited on the latest observation. The base action is generated from an older observation, but the edit policy transforms the remaining actions using the observation available at the instant of execution — so the policy stays maximally reactive despite the delay it was built around.
• Q-value chunk selection. At inference the VLA proposes candidate chunks; once the inference delay has elapsed, the edit policy edits the remaining actions and selects the chunk with the highest Q-value.

BibTeX

@misc{author2026method,
      title={Real-Time EXPO-FT: Reinforcement Learning for Real-Time Vision-Language-Action Policies},
      author={Author One and Author Two and Author Three and Author Four},
      year={2026},
      eprint={XXXX.XXXXX},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/XXXX.XXXXX},
}