When Does Adaptive Guidance Help? Belief-Aware Privileged Distillation for Autonomous Driving Under Partial Observability
Researchers present Belief-Aware GSAC, an adaptive knowledge distillation method for autonomous driving that modulates teacher guidance based on ensemble disagreement. Testing reveals that adaptive guidance helps under mild-to-moderate partial observability but fails under severe occlusion due to 'observability blindness'—where ensembles achieve low disagreement on visible data while missing occluded information.
This research addresses a fundamental challenge in autonomous driving: training agents that perform well with incomplete sensory information while leveraging knowledge from fully-informed teachers. The core innovation adapts the distillation coefficient dynamically based on how much an ensemble of models disagrees, theoretically allowing the student agent to rely more heavily on teacher guidance when uncertainty peaks. However, the empirical results expose a critical failure mode that has broader implications for uncertainty-aware machine learning systems.
The observability blindness phenomenon demonstrates why ensemble disagreement alone is insufficient for detecting what an agent cannot perceive. When the ensemble trains exclusively on partial observations (matching the student's actual sensory constraints), it naturally achieves low disagreement precisely in regions where data is missing—the very situations where adaptive guidance should intensify. This creates a paradox where the uncertainty metric behaves opposite to its intended purpose under severe conditions.
For the autonomous driving industry, this finding suggests that current ensemble-based uncertainty quantification methods may provide false confidence in partially-observable environments, potentially masking dangerous decision-making scenarios. The proposed architectural fix—training the ensemble on privileged full-state predictions—shows promise but remains unvalidated in this work. Meanwhile, the practical result that a simple linear decay schedule outperforms the adaptive approach raises questions about whether complex uncertainty mechanisms justify their computational overhead for this domain.
Future research should focus on designing uncertainty metrics that explicitly account for observability gaps rather than relying on prediction disagreement as a proxy. This work establishes ensemble prediction targets as a critical hyperparameter choice with downstream effects on safety-critical applications.
- →Ensemble disagreement-based adaptive guidance fails under severe partial observability due to observability blindness, where low disagreement masks missing information rather than indicating confidence
- →Simple linear decay schedules outperform adaptive distillation coefficients on challenging partial-observability tasks, suggesting scheduling effects may matter more than ensemble uncertainty
- →Training ensembles on privileged full-state predictions could address observability blindness but requires architectural changes and validation
- →Uncertainty quantification methods designed for complete observations may provide misleading signals in partially-observable autonomous driving scenarios
- →Ensemble prediction targets emerge as a critical design choice with significant implications for uncertainty-aware teacher-student learning frameworks