CycleVLA: Proactive VLA for Self-Correcting Robots

• CycleVLA is a proactive Vision–Language–Action system that integrates progress-aware action modeling with vision–language feedback to predict and prevent robotic execution failures.
• It employs a VLM-based failure predictor coupled with a cyclic backtracking mechanism and Minimum Bayes Risk decoding to select optimal corrective actions.
• Empirical results demonstrate significant improvements in long-horizon, under-trained task settings, enhancing robotic success rates by up to 6–10 percentage points.

CycleVLA is a Vision–Language–Action (VLA) system engineered to endow robot agents with proactive, self-correcting capabilities. Unlike traditional frameworks for robotic failure detection and correction that operate primarily in a post hoc fashion, CycleVLA anticipates potential execution failures and initiates recovery before faults fully manifest. The system systematically integrates progress-aware action modeling, vision–language feedback for failure forecasting and planning, and optimal hypothesis selection at test time. CycleVLA achieves significant improvements in long-horizon and under-trained task settings, contributing a robust approach to proactive failure correction in multi-modal sequential decision-making domains (Ma et al., 5 Jan 2026).

1. System Architecture and Foundations

CycleVLA unifies three principal modules into a cyclic control loop:

• Progress-aware VLA: Enhances the base action policy by explicitly predicting subtask completion progress and termination points, serving as an intrinsic alarm mechanism at transition bottlenecks.
• Vision-LLM (VLM)-Based Failure Predictor and Planner: Uses an off-the-shelf VLM (e.g., GPT-5.2) to predict impending failure and recommend backtracking to restore task preconditions.
• Minimum Bayes Risk (MBR) Decoding at Test Time: Leverages stochastic policy rollouts, selecting the predictive action hypothesis most central to the distribution for successful retry after failure anticipation.

CycleVLA operates in a closed feedback loop: The progress-aware VLA marks transition readiness; when flagged, the VLM evaluates imminent success or failure and, if necessary, triggers subtask backtracking using trajectory state reversal. Upon backtracking, the MBR decoding step samples and ranks alternative action chunks, selecting the one with the minimum average risk for the retry.

2. Progress-Aware Action Modeling

Subtask progress awareness is motivated by the observation that execution failures in robotic manipulation concentrate at subtask transitions (e.g., object grasp or insertion boundaries) (Ma et al., 5 Jan 2026). To capture this, the action output vector is extended:

• Action Output Augmentation: Original actions $a_t \in \mathbb{R}^7$ (cartesian deltas, orientation deltas, gripper command) are augmented to $a_t \in \mathbb{R}^9$ by adding:
  • Binary stop signal $s_t \in \{0,1\}$ (subtask termination indicator)
  • Progress estimate $p_t \in [0,1]$, quantized to 10 bins

Progress signals are supervised via L2 loss for deltas, binary cross-entropy for $s_t$, and regression or cross-entropy for $p_t$. Subtask labels and alignment are generated with LLM-based task decomposition and motion primitive segmentation. A "last-step oversampling" strategy emphasizes accurate subtask termination signals during fine-tuning.

A transition detection protocol defines a threshold $\tau_p$ (typically 0.9) for $p_t$, employing confirmation heuristics over multiple timesteps to filter spurious positives before invoking downstream modules.

3. Vision–Language Failure Prediction and Backtracking

The VLM-based failure predictor receives third-person and wrist-camera images, along with the current and global subtask context. Once progress exceeds $\tau_p$, a VLM is queried to answer:

1. Will the subtask succeed upon continuation?
2. If not, to which subtask index $j$ should the system backtrack?

When a backtrack is advised, recorded action deltas are reversed to the step prior to the indicated subtask, with the retry counter capped at $R$ (default 3). The state is reset, and new action candidates are generated for selection.

Formally, for state $o_t$ and target subtask $g_k$, the VLM $\mathcal{F}$ is queried with prompt $X_t$ and returns $(dec, j)$, where $dec$ denotes {transit, backtrack} and $j$ is an index. The conditional failure probability is estimated as $P(\text{failure} | o_t, g_k) \approx 1$ if $dec = \text{backtrack}$, 0 otherwise.

4. Minimum Bayes Risk Decoding Mechanism

To prevent repeated local policy errors after backtracking, CycleVLA employs test-time MBR decoding:

• Sampling: For each retry attempt, $N$ candidate action trajectories ($a^{(i)}_{t:t+H-1}$) are sampled from the stochastic VLA policy.
• Risk Computation: Each candidate $a^{(i)}$ is scored by mean L2 trajectory distance against all samples:

$$R(a^{(i)}) \approx \frac{1}{N} \sum_{j=1}^N d(a^{(i)}, a^{(j)})$$

where $d(\cdot, \cdot)$ is L2 distance over concatenated end-effector state sequences.

• Selection: The medoid (minimum average risk) hypothesis $a_\text{MBR}$ is executed.

Empirical ablation demonstrates that omitting MBR causes a substantial performance drop (success rate 95.3%→92.5%) and increases runtime (by 40%) (Ma et al., 5 Jan 2026).

5. Training Protocol and Test-Time Integration

CycleVLA builds on the OpenVLA diffusion policy backbone, fine-tuned on the full LIBERO benchmark task suite for 500k steps. Each training pass uses action chunks of length $H=8$ with LoRA adapters, resulting in approximately 313M parameters for proprioceptive and action heads. The MBR mechanism is not incorporated during training; it is a zero-shot test-time scaling strategy requiring no further knowledge distillation or fine-tuning.

At test time, each failed subtask is retried up to $R=3$ times with $N=8$ sample trajectories, applying MBR selection. This consistently improves both under-trained and fully trained agent success rates, yielding gains of +6–10 percentage points (see Table 2 below).

Checkpoint	SR w/o FC	SR w FC	Δ
200K	73.2	80.0	+6.8
350K	83.2	89.2	+6.0
500K	89.3	95.3	+6.0

SR: Success Rate; FC: Failure Correction (CycleVLA mechanism)

6. Empirical Evaluation and Ablation Studies

CycleVLA achieves state-of-the-art results on the LIBERO robotic manipulation benchmarks. Long-horizon success rates improve from baseline methods by approximately +18 percentage points and outperform other advanced VLA policies (e.g., GR00T N1, FPC-VLA, ThinkAct). Success rates remain robust across spatial, object, goal, and long suites, with aggregate performances as follows:

Method	Spatial	Object	Goal	Long	Avg
Diffusion Policy	78.3	82.5	68.3	50.5	72.4
OpenVLA	84.7	88.4	79.2	53.7	76.5
ThinkAct	88.3	91.4	87.1	70.9	84.4
FPC-VLA	87.0	92.0	86.2	82.2	86.9
GR00T N1	94.4	97.6	93.0	90.6	93.9
CycleVLA	97.6	98.1	91.7	93.6	95.3

Ablation experiments confirm the criticality of both the progress-aware signal and the MBR retry strategy. The removal of stop-signal oversampling or use of alternate (smaller) VLMs leads to marked declines in task completion. Exclusive reliance on failure cutoffs or absence of MBR disproportionately harms long-horizon task reliability. Runtime analysis on A10 GPUs indicates cycle overheads are dominated by VLA rollout (68%), with 0.1% attributable to MBR computations.

7. Relation to Temporal/Cyclic Reasoning and Future Directions

While CycleVLA addresses subtask-level proactive recovery in vision–language–action agents, the closely related CycliST benchmark (Kohaut et al., 30 Nov 2025) reveals that existing video-LLMs lack fundamental temporal and cyclic reasoning abilities. CycliST demonstrates substantial deficits in periodic pattern recognition, temporal quantification, and cycle-attribute binding across widely used VLM architectures. These results suggest CycleVLA’s explicit subtask progress–tracking and cycle-consistent action selection are partly addressing the lack of inductive temporal bias in modern VLMs.

Proposed architectural enhancements for future CycleVLA-like models include integrating explicit temporal modules, cycle-detection layers (e.g., via Fourier/spectral analysis), physics-inspired dynamical models, and graph-based spatiotemporal reasoning frameworks. Training could benefit from synthetic cyclical task augmentation, curriculum learning over increasing periodic complexity, and large-scale real-world cyclical scenario pretraining.

Current limitations include reliance on reversible dynamics for action backtracking, moderate test-time computational overhead, and dependence on external VLM oracles for failure prediction. Integration of end-to-end failure reasoning and real-robot validation in high-complexity, non-reversible domains constitute active research frontiers.

• CycleVLA: Proactive Self-Correcting Vision-Language-Action Models via Subtask Backtracking and Minimum Bayes Risk Decoding (Ma et al., 5 Jan 2026)
• CycliST: A Video LLM Benchmark for Reasoning on Cyclical State Transitions (Kohaut et al., 30 Nov 2025)