PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Chenyang Ma, Kai Lu, Guangyu Yang, Jiuming Liu, Shitong Xu, Bill Byrne, Ioannis Havoutis, Niki Trigoni, Andrew Markham

arXiv:2601.02295v2cs.RO

TL;DR

Robot failure correction is usually post hoc, leaving limited opportunity to recover once errors fully occur. CycleVLA combines progress-aware subtask execution, VLM-guided backtracking, and MBR retry selection to provide proactive self-correction. It improves success across simulation VLAs, reaches a 91% average real-robot success rate, and corrects approximately 80% of injected failures.

  • Problem

    Existing robot failure detection and correction is typically post hoc, while proactive self-correction during execution remains an open challenge.

  • Method

    CycleVLA combines subtask progress and stop prediction, VLM-based failure prediction and backtracking, and MBR decoding for retry trajectory selection.

  • Results

    91% average success rate on one precise and two long-horizon real-robot tasks, with approximately 80% of manually injected failures corrected.

  • Takeaways & Limitations

    CycleVLA improves success by correcting execution failures across VLAs of varying capability, and MBR provides zero-shot test-time scaling for VLA policies.

  • Takeaways & Limitations

    Backtracking assumes reversible robot state transitions, while MBR increases inference time through multiple forward passes.

Abstract

from arXiv · show

Current work on robot failure detection and correction typically operates 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 on the LIBERO and LIBERO-Plus simulation benchmarks show that CycleVLA surpasses the state-of-the-art VLA π0.5, improves success rates by correcting execution failures across VLAs of varying capability, from under-trained ones to fully converged policies, and that MBR serves as an effective zero-shot test-time scaling strategy for VLAs. On a real robot, CycleVLA reaches a 91% average success rate on one precise and two long-horizon manipulation tasks. We further conduct stress tests with multiple manually injected perturbations (e.g., swapping in a distractor at the expected location while relocating the true target object mid-execution), where CycleVLA corrects ~80% of injected failures and maintains success rates comparable to unperturbed execution. Project Page: https://dannymcy.github.io/cyclevla/

I. INTRODUCTION

CycleVLA equips VLAs with proactive self-correction by anticipating failures near subtask transitions, backtracking when failure is predicted, and improving retries with MBR decoding.

  • Motivation: CycleVLA targets proactive self-correction, detecting incipient execution errors and adapting before failure fully manifests rather than correcting failures post hoc.The system is designed to complete tasks within the execution episode without external intervention.
  • Method: The method adds explicit subtask progress and stop prediction by decomposing demonstrations into aligned subtasks and extending the VLA action dimensions.Progress cues help identify when subtasks are nearing completion, where failures frequently occur.
  • Method: A VLM predicts failures and plans either continued execution or backtracking to the earliest subtask that restores missing preconditions.This decision is made at test time near subtask termination.
  • Method: MBR decoding samples multiple VLA trajectories and selects a consensus candidate to improve retry success after backtracking.The approach is presented as a zero-shot test-time scaling strategy for VLAs.
  • Results: Extensive LIBERO and LIBERO-Plus experiments show improved success rates across VLAs of varying capability, including results surpassing π0.5.The reported gains include effectiveness on under-trained VLAs.
  • Results: On a real robot, CycleVLA reaches a 91% average success rate across one precise and two long-horizon tasks and corrects approximately 80% of injected failures.These stress tests used multiple manually injected perturbations.

III. PRELIMINARIES

CycleVLA equips VLAs with subtask-aware progress and stop prediction, VLM-guided backtracking, and MBR-based retry selection. Its inference procedure monitors progress, chooses whether to transition or backtrack, restores the target subtask, and retries with consensus action chunks.

  • Preliminaries: A VLA maps observations and language goals to continuous chunks of future robot actions, with stochastic decoding enabling multiple action hypotheses.Actions include translational and rotational end-effector displacements plus gripper state.
  • Learning Stop and Progress Signals: CycleVLA constructs subtask-decomposed demonstrations and extends VLA outputs from 7 to 9 dimensions with stop and progress signals.The stop signal marks subtask termination, while progress is discretized into 0.1 bins based on normalized within-subtask time.
  • Learning Stop and Progress Signals: Stop and progress are predicted as scalar outputs alongside end-effector actions, requiring only a widened action dimension rather than separate classification heads.The design distinguishes precise termination detection from approximate proximity-to-completion estimation.
  • Subtask Backtracking: At a progress threshold, a VLM decides whether to transition or backtrack to the earliest subtask that restores missing preconditions.Backtracking restores the robot configuration by reverse-executing recorded delta actions but does not reset object states.
  • MBR Retry Selection: After backtracking, MBR samples multiple action chunks and selects the hypothesis with minimum expected risk, approximated from pairwise distances among sampled trajectories.Distances use predicted end-effector motion, and consensus favors trajectories in high-density regions of the policy output space.
  • Overall Inference Procedure: The complete inference loop observes the robot, samples and executes action chunks, checks success, invokes VLM planning near subtask completion, and either retries or advances subtasks.The procedure includes configurable progress threshold, chunk size, sample count, retry limit, and timeout.

V. SIMULATION EXPERIMENTS AND ANALYSIS

The simulation experiments evaluate CycleVLA’s ability to improve task success and recover failed trajectories across VLAs with different training levels, while also measuring MBR’s contribution and each component’s impact.

  • Simulation Experiments: The experiments ask whether proactive self-correction improves success and recovers failures relative to a monolithic VLA, including with under-trained VLAs.They also examine MBR’s contribution to VLA inference gains and the impact of individual components.

A. Implementation Details

CycleVLA is evaluated on LIBERO and LIBERO-Plus using two VLA backbones, including clean, perturbed, and under-trained settings. Its failure correction improves long-horizon and perturbation robustness, though gains are limited when perturbations directly disrupt perception and control.

  • CycleVLA evaluates π0.5 and OpenVLA-OFT-Diff on LIBERO’s four suites and LIBERO-Plus, which probes seven perturbation dimensions.
  • CycleVLA improves average LIBERO performance by +4.5 over OpenVLA-OFT-Diff and +1.6 over π0.5, whose average success rate reaches 98.5.
  • CycleVLA improves LIBERO-Plus robustness by +7.4 over OpenVLA-OFT-Diff and +21.0 over π0.5 under unseen perturbations.
  • Camera and robot-state perturbations limit gains for the OpenVLA-OFT-Diff variant because they more directly affect perception and control.
  • CycleVLA provides consistent gains across 200K, 350K, and 500K checkpoints, with earlier corrected checkpoints approaching later uncorrected performance.

C. Analysis of MBR Decoding

The MBR analysis evaluates consensus selection over multiple stochastic action hypotheses and studies how sampling count and distance metric affect estimated success. MBR improves success across VLA capabilities, especially for under-trained models, while L2 performs best among tested metrics.

  • Evaluation protocol: MBR evaluates action-chunk selection from N stochastic hypotheses across 200 episodes per task and averages success over aligned decision steps and episodes.
  • Number of hypotheses: Increasing N generally improves MBR performance, with most gains from 4 to 8 hypotheses and a plateau beyond N = 16.
  • Baseline: Random selection approximates the base VLA’s marginal behavior because it samples action chunks directly from the stochastic policy distribution.
  • Results: MBR consistently improves success probability across VLA capabilities, with larger gains for 200K and 350K under-trained models than for the 500K model.
  • Distance metric: L2 achieves the best performance across tested VLAs, followed by L1, while cosine and correlation yield the smallest improvements.

D. Ablation Studies

Ablations show that MBR, VLM-guided backtracking, and explicit stop/progress signals each contribute to CycleVLA’s performance. The VLM-gated design preserves most of always-on MBR’s gain while avoiding its substantial inference-time cost.

  • Removing stop prediction and last-action oversampling degrades performance through spurious high-progress predictions and premature termination.
  • Removing MBR causes a moderate success-rate drop, while retrying up to R times still provides partial robustness.
  • Replacing GPT-5.2 with LLaMA-3.2-11B reduces success because it more often transits instead of backtracking, missing correction opportunities.
  • Always-on MBR achieves the highest success rate but requires approximately 2.2× the inference time of the main VLM-gated method.
  • Terminating when the VLM predicts failure causes a modest approximately 10% drop, whereas the full design uses predicted failure to trigger backtracking and retry.

B. Evaluation Protocols: Natural Failures and Injected Errors

The evaluation covers three real-world manipulation tasks under natural failures and human-injected errors. Stress tests place progressively harder perturbations near subtask completion to measure whether CycleVLA detects failures, backtracks, and recovers.

  • Tasks: The real-robot tasks are teapot hanging, fruit sorting, and cookware packing.The tasks include one precision task and two long-horizon tasks, with randomized object positions and orientations.
  • Natural Failures: Natural-failure evaluation compares CycleVLA with a baseline that transitions whenever the trained VLA predicts a stop signal.Failures arise only from autonomous policy execution in this setting.
  • Injected Errors: Injected errors occur near subtask completion, requiring the VLM to detect failure and backtrack rather than transition forward.The protocol tests distractor injection, target displacement, and target substitution with increasing difficulty.
  • Perturbation Types: The perturbations test distractor discrimination, recovery from changed placement targets, and recognition of displaced true targets replaced by unrelated objects.Target substitution combines a distractor at the expected location with relocation of the original target.
  • Results: CycleVLA outperforms the baseline across all three natural-failure tasks, with success rates of 0.87 vs. 0.73, 0.93 vs. 0.87, and 0.93 vs. 0.80.It corrects ∼80% of injected failures on average, while performance declines from distractor injection to target displacement to target substitution.
  • Limitations: The approach remains scope-limited by its assumption that robot state transitions are reversible and by the added inference time required for MBR decoding.These constraints are especially relevant to highly dynamic, irreversible, or contact-rich settings.

Appendix for CycleVLA

The appendix details how demonstrations are converted into movement primitives, gripper-segmented subtasks, and temporally aligned training data. It also describes thresholding, filtering, and downsampling procedures used before subtask-boundary inference.

  • Movement Primitive Extraction: Movement primitives classify translation, rotation, and gripper behavior from robot-state differences over a four-timestep sliding window.Components below threshold are omitted, and no detected movement is labeled stop.
  • Thresholding: LIBERO uses separate translation, rotation, and gripper thresholds of 0.02, 0.0075, and 0.03, respectively.The translation threshold is later optimized per trajectory to reduce movement–gripper overlaps and spurious stop labels.
  • Trajectory Segmentation: Trajectories are segmented at gripper-state transitions, with chunks labeled close, open, or idle according to their dominant gripper value.Three gripper thresholds are combined by voting, followed by filtering of isolated idle segments.
  • Subtask Boundary Inference: Trajectories longer than 100 steps are uniformly downsampled before LLM boundary inference, then mapped back to the original indices.The procedure preserves global temporal structure while limiting the sequence length presented to the LLM.

B.4. Qualitative Results of Subtask-Decomposed Dataset

The appendix evaluates subtask decomposition quality and documents implementation and runtime details for CycleVLA’s MBR-based test-time scaling. The system uses consensus over sampled trajectories while adding moderate overhead beyond standard action rollout.

  • Dataset Quality: Human evaluation compares LLM-predicted subtask boundaries with manually corrected timestamps across 40 demonstrations and five evaluators.Mean absolute timestamp deviation and relative boundary error are reported, with rotational and weak gripper boundaries identified as more challenging.
  • MBR Decoding: MBR samples N=8 stochastic action hypotheses and selects the hypothesis with minimum average pairwise L2 distance.The selected trajectory is the medoid under the trajectory-feature distance and is executed without modification.
  • Trajectory Features: Trajectory features accumulate translational and rotational deltas so that deviations compounding over a chunk influence the distance.This measures divergence in resulting motion rather than only step-wise action differences.
  • Signal Detection: Confirmation filters noisy stop and progress predictions before triggering subtask transitions or VLM checks.It accepts either two consecutive high signals or a recurring high signal after at least two low-signal steps.
  • Runtime: CycleVLA adds approximately 30% overhead over baseline runtime, while MBR selection itself contributes less than 0.1%.Standard action rollout remains the dominant runtime bottleneck across the analyzed GPU settings.

E. ADDITIONAL DETAILS OF REAL SYSTEM

The real system uses a 6-DOF robot arm with end-effector control, augmented action outputs, teleoperated demonstrations, and deterministic joint-state replay for backtracking. The three tasks use randomized initializations and real-world subtask annotations.

  • Task Initialization: Object positions are randomized within ±30 cm and orientations within ±60° for the three real-world tasks.The tasks shown are teapot hanging, fruit sorting, and cookware packing.
  • Action Representation: Each action contains translational and rotational displacements, a gripper command, a stop signal, and a subtask-progress signal.The policy outputs 9-dimensional action vectors and executes the first five actions of each chunk before requerying.
  • Demonstrations: Demonstrations are collected through leader–follower teleoperation, with operators marking subtask boundaries during recording.Each task uses 100 demonstrations captured with wrist and third-person RGB cameras.
  • Backtracking: Backtracking rewinds the arm by replaying recorded joint positions in reverse before resuming end-effector control for the retry.This implements the explicit subtask-level recovery mechanism on the physical robot.

F. PROMPT DETAILS OF SUBTASK DECOMPOSITION

The prompt details specify how CycleVLA decomposes task instructions and trajectories into minimal, exhaustive subtasks with timestamp boundaries. LLM prompts handle formal subtask generation, noisy labels, and trajectory alignment.

  • Timestamp Alignment: Subtask timestamps are inferred by aligning LLM-proposed subtasks with movement-primitive sequences and trajectory segments.The section presents exact prompts for proposing subtasks and inferring their timestamp boundaries.
  • Subtask Decomposition: The task-decomposition prompt asks for a minimal set of formal subtasks sufficient to complete the robotic task.It emphasizes object-centric spatial cues and precise robotic language.
  • Trajectory Labeling: The trajectory prompt maps every subtask to contiguous start and end step indices without gaps or skipped steps.The required output is a labeled dictionary assigning each trajectory step to exactly one subtask.
  • Noise Handling: The labeling rules treat intermediate stop labels and brief inconsistent movement or gripper labels as possible noise.Meaningful gripper actions are generally more reliable than motion primitives, and labels are cross-referenced with neighboring steps.

G. PROMPT DETAILS OF FAILURE PREDICTOR AND PLANNER

The failure-prediction prompts ask a VLM to forecast execution success at subtask boundaries and choose transit or backtrack using synchronized global and local visual evidence. Recovery targets the earliest subtask that restores the missing precondition, including in wrong-object or wrong-subtask cases.

  • Planner Output: The planner outputs the next exact subtask, a transit or backtrack decision, and a concise reason grounded in observable cues.It also records success likelihood, key risks, view agreement, and the decision basis.
  • Failure Forecasting: At approximately 90% subtask progress, the VLM forecasts whether continuing without corrective repositioning will succeed.The default is transit unless strong, unambiguous visual evidence predicts failure.
  • View Fusion: The VLM combines FRONT global context with WRIST local interaction cues to assess functional success.FRONT captures identity, alignment, reachability, and clearance; WRIST captures orientation, contact, slip, stability, and affordances.
  • Affordance Reasoning: The prompts evaluate whether gripper poses match object affordances, including bowl rims, drawer handles, and push or pull directions.Partial or asymmetric contact is not penalized when it provides a valid and stable affordance.
  • Recovery Planning: Wrong-object or wrong-subtask engagement triggers backtracking to the earliest step that restores correct target selection and preconditions.The prescribed target is typically reach, alignment, or identification rather than a trivial gripper action.
Loading 2601.02295v2…