Beyond Waypoint Regression: Query-Based Cost Learning over Reachable Ego Futures for End-to-End Driving

TL;DR

The paper introduces a query-based cost-learning approach for end-to-end driving that evaluates dynamic ego trajectory queries using compact scene tokens and cost aggregation. It improves safety by reducing collision rates while maintaining trajectory diversity and interpretability, outperforming prior methods on nuScenes and real-world logs without fine-tuning.

Why It Matters

This approach enables safer, more adaptable autonomous driving systems by directly incorporating safety constraints at runtime, avoiding reliance on rigid imitation learning. The method supports real-time planning with interpretable cost functions, making it suitable for deployment in unpredictable environments.

Key Findings

• The framework uses query-based cost estimation over reachable ego futures instead of dense grid or fixed trajectory outputs.
• Contingency-aware cost aggregation and MPPI mixing enable effective exploration of diverse, feasible trajectories.
• The method achieves lower collision rates than state-of-the-art planners on both benchmark and real-world data.