Belief-Aware GSAC Improves Autonomous Driving Under Partial Observability
A recent paper published on arXiv (2605.26155) presents Belief-Aware Guided Soft Actor-Critic (BA-GSAC), a technique that dynamically adjusts knowledge distillation from a comprehensive full-state teacher to a student with limited observations in autonomous driving scenarios. In contrast to traditional GSAC, which employs a constant distillation coefficient lambda, BA-GSAC modifies lambda according to ensemble disagreement, providing a framework to explore the effectiveness of adaptive guidance. Tests conducted on Highway-Env at three POMDP difficulty levels evaluated five approaches: fixed lambda (0.01, 0.1), adaptive, linear decay, and standard SAC. Initial single-seed results indicate advantages in mild and moderate partial observability, but under severe occlusion (assessed with 3 seeds for all methods), the adaptive coefficient drops to lambda_min within roughly 3,000 steps, due to an observability blindness effect.
Key facts
- BA-GSAC modulates distillation coefficient lambda via ensemble disagreement.
- Five strategies tested: fixed lambda (0.01, 0.1), adaptive, linear decay, vanilla SAC.
- Experiments conducted on Highway-Env across three POMDP difficulty levels.
- Under severe occlusion, adaptive coefficient collapses to lambda_min within ~3K steps.
- Observability blindness: ensemble predicts partial observations, leading to low disagreement despite uncertainty.
- Preliminary single-seed runs show benefits under mild and moderate partial observability.
- Study uses 3 seeds for all methods under severe occlusion.
Entities
Institutions
- arXiv