acceptodds
Desk rejected at ICLR 2027

Plan2Plan: Rethinking Trajectory Generation as Plan-to-Plan Transport for End-to-End Autonomous Driving

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

21 views

Abstract

Trajectory proposal-based planning has emerged as a promising paradigm for end-to-end autonomous driving. However, existing generative planners typically produce trajectory candidates from scratch using Gaussian noise or generic trajectory anchors. Such scratch-to-plan transport incurs a substantial transport gap, complicating trajectory generation and compromising plan quality. In this work, we propose Plan2Plan, a generative framework that reformulates candidate generation as a plan-to-plan transport problem, directly transforming preliminary plans from an off-the-shelf planner (termed expert priors) into high-quality driving trajectories. Specifically, we develop an expert prior-guided Shortcut flow-matching model that recasts the velocity field as a scene-conditioned corrective field over preliminary plans. Starting from structured, plausible plans substantially shortens the transport, shifting the generative task from complete trajectory construction to residual plan correction and enabling reliable single-step refinement. Crucially, to supervise this corrective flow matching without collapsing multimodal modes, we avoid uniformly pulling all expert priors toward the single ground-truth demonstration. Instead, we introduce a geometry-aware multi-expert supervision scheme that preserves strong priors while steering suboptimal ones toward their geometrically closest higher-quality trajectory, elevating plan quality while safeguarding driving intent. Extensive experiments on NAVSIM-v1, NAVSIM-v2, and closed-loop HUGSIM demonstrate that Plan2Plan outperforms strong generative planning baselines across diverse evaluation settings.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
End-to-End Autonomous Driving, Trajectory Generation

Related papers

Discussion (0)

Sign in to comment.