Beyond Waypoint Regression: Query-Based Cost Learning over Reachable Ego Futures for End-to-End Driving
Query-based cost learning over reachable ego futures improves end-to-end driving collision rates on nuScenes and real logs.
The paper replaces waypoint regression with a query-based cost learner that scores dynamically reachable ego trajectories. Compact joint scene tokens, contingency-aware cost aggregation, and cost-guided intra-cluster MPPI mixing turn the cost map into feasible plans. On nuScenes it improves collision rate over ST-P3 and NMP while remaining competitive on L2. On real driving logs it reduces collisions versus SparseDrive and Alpamayo without fine-tuning.
- Learns bounded costs for reachable ego futures instead of waypoint regression
- Beats ST-P3 and NMP on nuScenes collision rate while staying competitive on L2
- Cuts collisions versus SparseDrive and Alpamayo on real logs without fine-tuning
- Retains an interpretable cost interface and diverse candidate trajectories
From the original article
1 min read at arxiv.orgEnd-to-end planners based on waypoint regression achieve strong open-loop accuracy, but they primarily learn to mimic expert geometry and remain difficult to adapt to deployment-time safety constraints. We propose a query-based cost-learning framework that estimates bounded costs for dynamically reachable ego trajectory queries, rather than dense BEV cells or a small regressed trajectory set.
By Ahmed Abouelazm · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.