CARE Teaches VLA Policies to Recover from Execution Failures

Introduction

Vision-language-action (VLA) policies are helping robots connect visual observations, natural-language goals, and physical actions. Yet strong performance on nominal trajectories does not necessarily translate into reliable behavior in the real world. A grasp may be slightly misaligned, an object may be left in an unexpected pose, or a transition between subtasks may fail. Once the execution state diverges from the training distribution, a policy can continue producing inappropriate actions or lose the entire task.

The paper CARE: Experience-Guided Atomic Corrective Execution for Vision-Language-Action Policies addresses this gap through a focused question: how can a robot recover from a local failure while preserving as much task progress as possible?

The core approach

• Learn from failures encountered during execution. Rather than depending primarily on hand-designed or random perturbations, CARE collects failed rollouts produced during operation. These failures are treated as useful experience for building recovery data.
• Model deviations by execution stage. A deviation during grasping is not equivalent to one during transport or placement. CARE therefore represents post-failure deviations conditionally on the current stage, then uses the empirical distributions to synthesize representative failure states and corrective demonstrations.
• Use atomic corrective execution. At inference time, the task is handled through stage-wise planning. When a mismatch is detected, the system can perform a small adjustment or re-operate a relevant step instead of restarting the complete procedure.
• Ground monitoring in 3D and physical state. The monitoring component tracks the scene and execution state in three dimensions, allowing corrective behavior to be triggered while the remaining task structure is still usable.

A benchmark centered on recovery

CARE also introduces the Failure State Recovery Benchmark, or FSR-Bench. It evaluates whether a policy can recover from intermediate failure states involving local deviations and structural anomalies. This focus extends beyond a single end-to-end success score. A robot may fail to complete an episode even when it has the capability to recover from the particular mistake; conversely, a high nominal success rate may hide serious brittleness once conditions change.

According to the paper’s abstract, CARE was tested with multiple VLA backbones, simulation benchmarks, and real-world dual-arm tasks. The reported average task-success gains are 14.5 percentage points in simulation and 15.9 points in the real world. The provided material does not break these results down by individual backbone or task, so they should be read as aggregate findings rather than a guarantee for every setting.

Why it matters

The main contribution is a data loop for recovery. Failures are not treated solely as discarded episodes; they become evidence about where execution drifts and what a useful correction should look like. The atomic-action perspective is particularly relevant to long-horizon manipulation, where restarting after every small error is inefficient and can introduce new errors.

The approach also leaves open practical questions. Its performance depends on the quality and coverage of failed rollouts, the reliability of stage identification, and the accuracy of 3D state monitoring. Generalization to more open-ended scenes and unseen object configurations will require further study. Still, the release of code, models, and data makes CARE a useful basis for reproducible research on failure-aware VLA systems.

Source: Hugging Face Daily Papers