SAFECAST: Robust Failure Detection for VLA Policies with Contrast-Set Training and Calibration

Vision-language-action (VLA) policies often struggle when deployed in environments that differ from their training data, such as those containing new objects, clutter, or varied lighting. Existing methods for detecting these failures, such as the ScAlable Failure Estimation (SAFE) framework, rely on hidden-state probes calibrated on source-distribution data. This paper investigates whether incorporating contrast-set perturbations—small, controlled changes to visual or language inputs—can improve the robustness and reliability of these failure detectors when the robot encounters out-of-distribution conditions.

The authors introduce SAFECAST, a framework that augments the training and calibration phases of failure probes. Instead of relying solely on source-distribution rollouts, SAFECAST generates contrast-set trajectories by applying visual perturbations (e.g., adding distractor objects) and language perturbations (e.g., paraphrasing instructions) to the robot's environment. These augmented trajectories are used to train the failure-prediction MLP and to calibrate the time-varying risk thresholds via functional conformal prediction. The authors evaluate this approach across both simulated (LIBERO) and real-world (DROID) robotic manipulation tasks using multiple VLM backbones like OpenVLA and π0.

SAFECAST consistently outperforms the baseline SAFE method in detecting failures under distribution shifts. The results show that:

• Combining both visual and language contrast sets yields the most robust failure detection.
• Augmented probe training and calibration significantly improve ROC-AUC and F1 scores across various risk tolerance levels (α).
• Sim-to-real transfer is improved; training probes in simulation with contrast-set augmentation and calibrating them on physical rollouts is more effective than relying on expensive, purely real-world data collection.