Autonomous driving AI policies lack timely prediction in critical scenarios

A new research paper explores the internal reasoning of autonomous driving policies, specifically focusing on prediction and planning capabilities. Researchers used probing techniques and targeted perturbations to analyze how these capabilities emerge with increased scale in both imitation learning and reinforcement learning models. The study found that despite strong performance in simulations, many policies struggle to form timely predictions of surrounding vehicles during critical near-collision events, indicating a limitation in the predictive signals used for ego planning. However, causal interventions demonstrated that correcting mistaken predictions can lead to safer ego planning trajectories. AI

IMPACT Reveals limitations in current autonomous driving AI, suggesting a need for improved predictive capabilities for safer navigation.

RANK_REASON Research paper published on arXiv detailing findings about AI model capabilities. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• arXiv
• Hyeonchang Jeon

• paper
• safety
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →