Answers are grounded in this paper and cite their sources.
End-to-end motion planning for autonomous vehicles is a critical research frontier, yet prevailing imitation learning approaches are fundamentally constrained by their reliance on regressing to a single, often sub-optimal, ground-truth trajectory. This paradigm fails to address the inherent multi-modality of driving scenarios and the interactive nature of traffic, limiting the system’s ability to discover safer or more efficient behaviors. To address these limitations, we propose a novel “generate-and-evaluate” framework for end-to-end motion planning that moves beyond direct imitation. The framework uses the sparse perception and multi-modal candidate-generation foundation of SparseDrive. Its ego-centric prediction module incorporates observed ego-vehicle, surrounding-agent, map, and temporal context. Subsequently, a principled trajectory selection methodology evaluates these candidates by constructing a predictive spatiotemporal risk field and utilizing a holistic cost function that synthesizes safety, efficiency, and comfort. A core innovation is a time-weighted safety metric that enables a more nuanced and proactive risk assessment. Experimental results show lower collision rate and longitudinal acceleration RMS on mini_val than SparseDrive, with nearly unchanged aver-age speed. On val, higher average speed and lower acceleration RMS come with a higher collision rate, indicating a trade-off among the evaluated criteria.
No takes yet. Share an insight, caveat, or question.
Liu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: