When Does Adaptive Guidance Help? Belief-Aware Privileged Distillation for Autonomous Driving Under Partial Observability
The paper introduces Belief-Aware GSAC (BA-GSAC), which adapts the distillation coefficient in autonomous driving under partial observability. The study finds that adaptive guidance is beneficial under mild and moderate conditions but fails under severe occlusion. The authors propose architectural fixes and highlight the importance of scheduling effects for stability in performance.
- ▪BA-GSAC modulates the distillation coefficient based on ensemble disagreement.
- ▪Preliminary results indicate that adaptive guidance helps in mild and moderate partial observability scenarios.
- ▪Under severe occlusion, the adaptive coefficient collapses, leading to poor performance.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Robotics arXiv:2605.26155 (cs) [Submitted on 24 May 2026] Title:When Does Adaptive Guidance Help? Belief-Aware Privileged Distillation for Autonomous Driving Under Partial Observability Authors:Mehmet Haklidir View a PDF of the paper titled When Does Adaptive Guidance Help? Belief-Aware Privileged Distillation for Autonomous Driving Under Partial Observability, by Mehmet Haklidir View PDF HTML (experimental) Abstract:Guided Soft Actor-Critic (GSAC) distills knowledge from a privileged full-state teacher to a partial-observation student for autonomous driving, but uses a fixed distillation coefficient lambda regardless of the agent's uncertainty.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
