Reproduce “Addressing Limitations of State-Aware Imitation Learning for Autonomous Driving” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Luca Cultrera, Federico Becattini, Lorenzo Seidenari, Pietro Pala, Alberto Del Bimbo
A multi-task learning agent for autonomous driving, built on a hierarchical Vision Transformer, addresses the inertia problem and the offline/online performance gap in state-aware imitation learning by propagating an enriched state token and employing noise injection. This approach reduces inertia-related failures by up to 32 percentage points and improves the correlation between offline and online metrics from -0.64 to -0.92.
Key insights
- Explicitly predicting stop/go intentions in an initial processing stage effectively mitigates the inertia problem by guiding the agent's decision-making on resuming movement.
- Augmenting the state token with noise significantly improves the alignment between offline evaluation metrics and online driving success, leading to more reliable model assessment.
- Explainability tools, leveraging attention mechanisms and ex-post semantic analysis, can differentiate between true causal confusion and visual 'hallucinations' as causes for apparent inertia failures, aiding targeted model improvements.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.