cs.ROAug 27, 2026

FLARE: A Failure-Aware Framework for Autonomous Correction and Recovery in Visual-Language Robotic Manipulation

Authors: Ganlong Zhao, Zijia Tang, Xingping Chen, Zhanghui Kuang, Ye Tian, Guanbin Li

Organizations: The Chinese University of Hong Kong · Centre for Perceptual and Interactive Intelligence · Duke University · Sun Yat-sen University · TengenX · Tencent Robotics X · Shenzhen Loop Area Institute · Guangdong Key Laboratory of Big Data Analysis and Processing

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Vision-Language-Action Models~(VLAs) have demonstrated significant promise in generalizing to complex, long-horizon robotic manipulation tasks. However, their performance remains brittle, as they are typically trained on trajectory-monotonic, failure-free demonstrations. This reliance on perfect" data leaves them unable to recover from common execution errors, such as a missed grasp, a dropped object, or an unexpected collision. In this paper, we propose FLARE, a novel framework that endows VLAs with robust error recovery capabilities through a Retry" and Reset" paradigm. First, we introduce a Retry" mechanism by injecting perturbation and bridging segments that decouple robot pose from environment state into demonstrations, enabling the policy to autonomously handle execution deviations. Second, to address critical, state-breaking (OOD) failures, we introduce a Reset" pipeline. We leverage an MLLM for offline failure analysis to automatically identify OOD states from execution videos. This analysis enables the efficient, targeted collection of a small library of object-centric Reset" skills, which are trained to restore the environment to a task-valid state. Our full framework integrates these learned policies. At inference, an online MLLM monitor arbitrates between task execution and ``Reset" skills. Experiments on challenging, contact-rich manipulation tasks show our approach significantly improves task success and robustness.

Figures & tables

Figure 1: FLARE: Failure-Aware Resilience in VLA. Previous methods are brittle, failing from minor perturbations (ID errors) or catastrophic states (OOD errors), and do not leverage the VLA’s inherent generative capability for continuous self-correction. Our FLARE framework introduces an ID/OOD error taxonomy to guide a dual Retry/Reset paradigm. We tackle ID errors via built-in Retry robustness and resolve OOD errors with MLLM-bootstrapped Reset skills, achieving robust, resilient autonomy.
Figure 2: The overall framework of our method. We first collect the failure data with the VLA model trained with regular demonstrations. Then we perform the failure analysis with MLLM and formulate the failures with ID/OOD taxonomy. For ID errors, we perform the perturbation & bridging augmentation to enhance the robustness and “retry” ability of VLA models. For OOD errors, we collect the object-centric reset skill data. All these data are utilized to obtained a skill bank for further close-loop inference.
MethodsCoffee_D0Coffee_D1Stack_D0Stack_D1StackThree_D0StackThree_D1Threading_D0ThreePiece Assembly_D0ThreePiece Assembly_D1Mean
OpenVLA [ 18 ]42%18%84%86%36%20%20%28%8%38.0%
Task-conditioned66%24%88%68%30%6%74%20%0%41.8%
Subgoal-conditioned76%26%88%74%24%6%78%20%2%43.8%
Motion-conditioned68%32%92%84%38%16%58%30%4%46.9%
Subgoal Self-reflection80%32%88%78%32%6%80%34%2%48.0%
Phoenix [ 38 ]94%48%96%86%50%20%68%52%6%57.8%
Table 1: Comparison of experimental results across 9 manipulation tasks in RoboMimic Simulation. The ‘D’ suffix in the task names denotes the range of object randomization in scene initialization. The results demonstrate that our failure-aware VLA framework for robust robotic manipulation achieves better performance than all previous methods as well as the strong VLA baselines. Phoenix-Human is gray as it involves human instruction as the self-reflection subgoal.
Real-World Task (40 trials)π0.5\pi_{0.5} (Baseline)Ours (FLARE)
Stack Three Blocks62.5%75.0%
Insert U-shaped Block45.0%55.0%
Table 2: Success rates of real-world manipulation tasks. Our method outperforms the baseline in two challenging settings.
Figure 3: Real-world experimental setup and execution trajectories. Top: The “Stack Three Blocks” task. Bottom: The “Insert U-shaped Block” task.
Figure 4: The sensitivity of two hyper-parameters, rotation rr and translation tt for perturbation & bridging. Our method achieves best performance when r=30∘r=30^{\circ} and t=0.7t=0.7 , respectively.
Figure 5: The failure case for Coffee and ThreePieceAssembly respectively. (a) Reset coffee pod; (b) Reset coffee machine lid; (c) Reset the T-shaped block; (d) Reset the U-shaped block.
MethodsCoffee_D0Coffee_D1ThreePiece Assembly_D0ThreePiece Assembly_D1
Ours w/o Reset92%74%60%54%
Ours Reset-Only88%64%60%50%
Ours96%78%62%58%
Ours-Oracle100%90%68%64%
Table 3: The performance comparison for our method and two variants, Ours w/o Reset and Ours-Oracle. Ours w/o Reset only applies the perturbation & bridging strategy for VLA model training, and Ours-Oracle replaces the generated instruction for object reset with human feedback. The experimental results demonstrate the effectiveness of our reset skill learning and error correction.
 Reset Skill Success RateGeneration Efficiency
TaskCoffeeThreePiece AssemblyCoffeeThreePiece Assembly
Reset Object 184%88%83.7%48.6%
Reset Object 224%20%11.6%5.9%
Table 4: The reset skill success rate and demonstration generation efficiency on two manipulation tasks. The Reset Object 1 is coffee machine lid/T-shaped block, and the reset object 2 is coffee pod/U-shaped block for Coffee/ThreePieceAssembly. Our method successfully generates demonstrations and trains useful VLA models that can reset the objects when errors occur.
TaskReset/RetryReset ObjectTimestamp
Coffee88%88%78%
ThreePiece Assembly96%78%66%
Table 5: The performance of failure case indentification of Gemini-2.5-Pro on three tasks: Reset/Retry classification, Reset object identification and Timestamp identification for Coffee and ThreePieceAssembly. The effective failure analysis provides high-quality reset dataset for further demonstration collection.
Figure 6: Failure cases for all tasks (stack task excluded as reset is not applicable). (a) Reset coffee pod; (b) Reset coffee machine lid; (c) Reset the needle-hole block; (d) Reset the T-shaped block; (e) Reset the U-shaped block; (f) Reset the blue block.
Figure 7: Visualization of a successful task execution for ThreePieceAssembly. Frame 1-4: The robot operates the regular task to assembly the three pieces by first pick-and-place the T-shaped block. Frame 5-8: The T-shaped block is not place appropriately and robot tries to reset it. Frame 9-16: The robot executes the regular task again and succeeds.
Table 6: The main prompt used in our experiment for video analysis.
Table 7: The main prompt used in our experiment for failure case detection during the closed-loop inference for ThreePieceAssembly.

Explore similar work

CardsList

  1. FailSafe: Reasoning and Recovery from Failures in Vision-Language-Action Models

    Oct 2, 2025Zijun Lin, Jiafei Duan, Haoquan Fang +4Robot Failure RecoverySynthetic Data Generation for Robotics

    1Nanyang Technological University · Centre for Frontier AI Research, A*STAR · 3Allen Institute for AI +1

  2. CARE: Experience-Guided Atomic Corrective Execution for Vision-Language-Action Policies

    Sep 21, 2026Junlan Xiao, Junwei Jiang, Zaibin Zhang +4Long-Horizon Robotic ManipulationVision-Language-Action Models

    Dalian University of Technology, Dalian, China

  3. RePO-VLA: Recovery-Driven Policy Optimization for Vision-Language-Action Models

    May 10, 2026Weijia Liufu, Xiaoyu Guo, Ruiyi Chen +16VLM RobustnessRobot Failure Recovery

    1Sun Yat-sen University · 2South China University of Technology · 3Peng Cheng Laboratory +3