Anomaly-Informed Confidence Calibration for Vision-Based Safety Prediction

Technical Summary: Anomaly-Informed Confidence Calibration for Vision-Based Safety Prediction

Problem Statement
Reliable confidence estimates are critical for deploying vision-based controllers in safety-critical autonomous systems, such as autonomous racing. Modern predictors often exhibit dangerous overconfidence when encountering test-time distribution shifts. A primary obstacle identified in this work is the "perception-dynamics gap": standard anomaly signals (e.g., autoencoder reconstruction error) operate solely in observation space. They effectively detect visual corruptions (e.g., darkness, blur) but fail to detect dynamics anomalies (e.g., actuation bias, control latency, unmodeled friction) where the image stream remains plausible while the vehicle's trajectory silently diverges. Existing post-hoc calibration methods, such as standard temperature scaling, rely on a single fixed scalar and degrade significantly under distribution shifts. Conversely, richer uncertainty quantification techniques like Deep Ensembles or MC Dropout incur computational costs incompatible with real-time control and treat all anomalies as a monolithic source of uncertainty.

Methodology
The authors propose an Anomaly-Informed Online Calibration framework that adjusts safety confidence without retraining the base safety predictor. The approach integrates signals from a frozen "world model" and test-time augmentation (TTA) into a lightweight calibrator. The pipeline consists of four stages:

1. Anomaly Score Extraction: The framework extracts two complementary, continuous anomaly scores from a learned world model (a VAE with latent dynamics):

   • Perceptual Score ($\rho_i$): Derived from the mean per-pixel reconstruction error of the VAE decoder over a context window. This captures visual distribution shifts.
   • Dynamics Score ($\delta_i$): Derived from the Mahalanobis distance of a 12-dimensional feature vector in the state-transition space. This vector includes Monte Carlo Dropout uncertainty estimates, action pattern statistics (e.g., freeze rates, jerk), and temporal correlations. This score captures anomalies in the transition dynamics that leave images visually nominal.
     Both scores are normalized to $[0, 1]$ using in-distribution statistics.

2. Test-Time Augmentation (TTA): To enhance visual robustness, the safety predictor $g$ is evaluated on the original image plus $M=8$ augmented versions (random photometric jitter and Gaussian noise). The predictions are averaged to produce a smoothed base prediction $\bar{p}_{i+k}$. This step reduces sensitivity to minor visual perturbations.

3. Anomaly-Conditioned Temperature Scaling: The core innovation is a dynamic temperature scaling mechanism where the effective temperature $T_{eff}$ is a function of the anomaly scores:
   $$T_{eff} = T_k \cdot \exp(w_\rho \rho_i + w_\delta \delta_i)$$
   Here, $T_k$ is a base temperature per prediction horizon, and $w_\rho, w_\delta$ are learnable weights. As anomaly evidence ($\rho_i, \delta_i$) increases, $T_{eff}$ increases, pulling the predicted probability toward 0.5 and reducing overconfidence. When anomalies are absent, the method reduces to standard temperature scaling.

4. Training Regime: The calibrator is trained on a mixture of in-distribution data and synthetically augmented data (using corruption types disjoint from the test protocols). The base safety predictor and world model remain frozen during calibration training.

Key Contributions

1. Online Calibration Framework: A method that fuses anomaly evidence from a world model with TTA to adjust safety confidence in real-time without retraining the base predictor.
2. Disentangled Anomaly Scoring: A strategy that separates perceptual and dynamics scores, addressing the perception-dynamics gap left by purely visual anomaly signals.
3. Physical Validation: A validation on a physical DonkeyCar platform demonstrating that calibration trained on synthetic augmentations transfers to qualitatively different real-world corruptions.

Experimental Results
The method was evaluated on a physical DonkeyCar under four unseen real-world anomaly protocols: Darkness (visual), Blur (visual), Actuation Bias (dynamics), and Processing Latency (dynamics).

• OOD Detection: The fused anomaly scores achieved a mean AUROC of 0.830 for distinguishing in-distribution from OOD data, significantly outperforming baselines like Feature Norm (0.658) and prediction-derived heuristics (approx. 0.54). Notably, the dynamics score provided substantial gains in detecting actuation bias and latency, where image-level methods struggled.
• Calibration Performance: The proposed method reduced the average Expected Calibration Error (ECE) from 0.184 (best baseline, DAC) to 0.116, representing a 37% improvement.
• Ablation Studies:
  • The perceptual score ($\rho$) provided the dominant calibration signal, reducing OOD-average ECE by 28% when added to TTA.
  • The dynamics score ($\delta$) contributed complementary robustness, particularly in reducing cross-seed variance and improving detection of dynamics-specific anomalies, though it did not significantly lower the mean ECE further when added to TTA+$\rho$.
  • The method preserved in-distribution performance, achieving a mean ECE of 0.121 on the held-out nominal test set.

Significance and Claims
The paper claims that reliable confidence estimation in safety-critical robotics requires moving beyond single-scalar calibration and purely visual anomaly detection. By explicitly modeling the "perception-dynamics gap" and conditioning the calibration temperature on disentangled anomaly evidence, the system can selectively reduce overconfidence under distribution shifts while maintaining nominal performance. The work demonstrates that synthetic augmentations can serve as a structural proxy for real-world corruptions, enabling a calibrator trained on synthetic data to generalize to unseen physical anomalies (e.g., hardware degradation) without retraining the underlying safety predictor. The authors note that while the method shows strong transferability, extrapolation beyond the training severity range remains a limitation.