End-to-end trajectory planners need to represent multiple plausible driving behaviors while producing a single executable trajectory under real-time constraints. Proposal-based approaches address this ambiguity by generating multiple candidates, but converting the proposal set in…
Safe autonomous driving requires both rapid responses to common high-risk events and deeper reasoning over rare, extreme long-tail scenarios in traffic safety. These scenarios are severely under-represented in naturalistic driving data, and existing trajectory and language-augmen…