In recent years, Flow Matching has emerged as a prominent method for generative modeling in robot motion generation. In its generic form, Flow Matching is an ODE-based neural sampler trained by regressing empirical flow fields associated with motion samples.
However, in robot motion generation, we often face additional constraints not considered in the collected data. Most current approaches train the flow on available data and use inference-time guidance to enforce task-specific constraints.
To address this mismatch, we propose ConFlow, a constraint-guided flow matching framework that incorporates constraint information directly into the training objective via differentiable barrier or cost functions. To meet design specifications such as smoothness and boundary conditions, we suggest replacing the standard Gaussian source distribution used in flow matching training with a conditional Gaussian Process.
Our approach also employs infeasible demonstrations as negative supervision, enhancing constraint satisfaction without the need for additional expert data. Experiments on two robot navigation tasks show that ConFlow achieves lower collision rates and higher trajectory quality than standard flow matching baselines, with or without inference-time guidance. These results validate the integration of training-time constraints as an effective method to bridge the training-inference gap in generative motion models.
Blogger's Review: ConFlow significantly enhances the quality and safety of robot motion generation by integrating constraint information directly into the training process. This innovative approach not only increases the applicability of flow matching but also provides new insights for the design of future intelligent robots. Its strategy for handling infeasible examples is particularly noteworthy, demonstrating how to leverage existing data to boost model robustness.