ConFlow: Constraints-Guided Learning with Flow Matching for Motion Generation

arXiv cs.AI · 6 hours ago

ConFlow is a constraint-guided flow matching framework that incorporates task constraints directly into the training objective for robot motion generation, rather than enforcing them only at inference time. The method replaces standard Gaussian source distributions with conditional Gaussian Processes and uses infeasible demonstrations as negative supervision to improve constraint satisfaction. Experiments on two-robot navigation tasks show ConFlow achieves lower collision rates and higher trajectory quality compared to standard flow matching baselines.