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

TL;DR

CycleVLA addresses the limitation of reactive robot recovery by enabling proactive self-correction in Vision-Language-Action systems. It combines a progress-aware VLA with a VLM based failure predictor and planner to decide when to transit or backtrack at subtask boundaries, and uses Minimum Bayes Risk decoding for robust test-time retries after backtracking. The approach improves task success on LIBERO benchmarks, particularly for long-horizon and under-trained VLAs, and demonstrates that MBR provides effective zero-shot test-time scaling. This work lays groundwork for integrating proactive failure reasoning into VLAs and suggests directions for end-to-end learned recovery and real-robot validation.

Abstract

Current work on robot failure detection and correction typically operate in a post hoc manner, analyzing errors and applying corrections only after failures occur. This work introduces CycleVLA, a system that equips Vision-Language-Action models (VLAs) with proactive self-correction, the capability to anticipate incipient failures and recover before they fully manifest during execution. CycleVLA achieves this by integrating a progress-aware VLA that flags critical subtask transition points where failures most frequently occur, a VLM-based failure predictor and planner that triggers subtask backtracking upon predicted failure, and a test-time scaling strategy based on Minimum Bayes Risk (MBR) decoding to improve retry success after backtracking. Extensive experiments show that CycleVLA improves performance for both well-trained and under-trained VLAs, and that MBR serves as an effective zero-shot test-time scaling strategy for VLAs. Project Page: https://dannymcy.github.io/cyclevla/

Figures (6)

• Figure 1: Vision–Language–Action models with proactive self-correction capabilities. We introduce CycleVLA, which enables VLAs to anticipate incipient failures and recover before execution collapses. CycleVLA first augments a VLA to estimate subtask-level progress and flag critical subtask transition points, where failures most frequently occur. Then, at these points during inference, a VLM is queried to predict whether the current execution will fail and to decide whether to backtrack. Upon backtracking, the VLA retries using test-time scaling via Minimum Bayes Risk decoding to improve success. This cycle repeats until the task succeeds or execution terminates.
• Figure 2: Pipeline for constructing the subtask-decomposed dataset. Following LLM subtask decomposition and extraction of movement primitives and gripper state segments, subtasks are directly aligned to gripper state segment timestamps when their counts match; otherwise, an LLM infers subtask boundaries from the movement primitive sequence. Please see Appendices \ref{['sec:Details and Evaluation of Subtask Decomposition']} and \ref{['sec:Prompt Details of Subtask Decomposition']} for more details.
• Figure 3: CycleVLA. (a) A finetuning pipeline that equips a VLA with subtask-level stop and progress prediction via extended action expert dimension and augmented subtask-decomposed training data. (b) At inference, predicted progress triggers a VLM-based failure predictor and planner, which decides whether to transit to the next subtask or backtrack, and selects the subtask to backtrack to. (c) After backtracking, the VLA retries execution using test-time scaling via MBR decoding to improve success. Please see Appendices \ref{['sec:Implementation Details of MBR Decoding']} and \ref{['sec:Prompt Details of Failure Predictor and Planner']} for MBR implementation details and VLM exact prompts.
• Figure 4: Qualitative examples of CycleVLA. CycleVLA performs multiple cycles of failure prediction, backtracking, and retry within a single long-horizon task, correcting errors across subtasks and achieving successful completion. More examples can be found in Appendix \ref{['sec:Additional Experiments and Details']}.
• Figure 5: Qualitative examples of subtask-decomposed dataset.