CycleVLA: Proactive Self-Correcting Vision-Language-Action Models via Subtask Backtracking and Minimum Bayes Risk Decoding

In this work, we introduce CycleVLA, a system that enables VLAs to anticipate incipient failures and recover before execution collapses.

Overview of CycleVLA.

Project website: https://dannymcy.github.io/cyclevla/

Paper: https://arxiv.org/abs/2601.02295

Summary video: https://www.youtube.com/watch?v=09W81JMbF1E

System Requirements

Training:

• 4 or 8 A100 GPUs with 40 GB VRAM

Inference:

• 1 GPU with ~24 GB VRAM

Installation

CycleVLA supports two VLA backbones, OpenVLA-OFT and pi0.5. See SETUP.md for instructions on setting up the environments. See LIBERO.md for setting up the LIBERO and LIBERO-Plus simulation benchmark and generating our subtask-decomposed dataset.

Training and Evaluation

See OPENVLA.md for finetuning/evaluating the OpenVLA-OFT backbone on the LIBERO and LIBERO-Plus simulation benchmark task suites.

See PI.md for finetuning/evaluating the pi0.5 (openpi) backbone on the LIBERO and LIBERO-Plus simulation benchmark task suites.

Real-Robot Experiments

If you also use the AgileX PiPER arm with a leader arm follower arm setup (more details in the paper), you can use our real robot code at https://github.com/dannymcy/cyclevla_distal/tree/main. The original codebase is from https://github.com/reeceomahoney/distal.

Support

If you run into any issues, please open a new GitHub issue.

Citation

@article{ma2026cyclevla,
  title={CycleVLA: Proactive Self-Correcting Vision-Language-Action Models via Subtask Backtracking and Minimum Bayes Risk Decoding},
  author={Ma, Chenyang and Lu, Kai and Yang, Guangyu and Liu, Jiuming and Xu, Shitong and Byrne, Bill and Havoutis, Ioannis and Trigoni, Niki and Markham, Andrew},
  journal={arXiv preprint arXiv:2601.02295},
  year={2026}
}