Unified Prediction and Planning via Conflict-Aware Disjoint Parameter Training
*, Seonae Jeon*, Giwon Lee*, Kuk-Jin Yoon†, Daehee Park† (* Equal contribution, † Corresponding authors)
European Conference on Computer Vision (ECCV) 2026
Abstract
This work identifies Skill Conflict in compact unified models for motion prediction and planning and proposes Disjoint Parameter Training (DPT) to separate task-critical parameter regions. Combined with sparse model merging, DPT improves prediction accuracy and planning safety on crowd navigation benchmarks while maintaining a compact model.