Generative Predictive Control: Flow Matching Policies for Dynamic, Difficult-To-Demonstrate Tasks
Vincent Kurtz, Joel Burdick
AI summary
Problem
Existing generative control methods rely on costly expert demonstrations and struggle with fast, nonlinear dynamics, limiting them to slow, quasi-static tasks.
Approach
The authors propose GPC, a supervised learning framework that alternates between generating training data via sampling-based predictive control and training a flow-matching policy, which is then used to warm-start control at inference.
Key results
- Outperforms PPO across multiple dynamic simulation tasks
- Novel warm-start scheme enables smooth high-frequency feedback
- Integrates risk-aware domain randomization for robust training
- Identifies scalability limits on complex humanoid tasks
Why it matters
Offers a scalable, supervised alternative to behavior cloning and reinforcement learning for fast, dynamic robotics applications.
Abstract
Generative control policies have recently unlocked major progress in robotics. These methods produce action sequences via diffusion or flow matching, with training data provided by demonstrations. But existing methods come with two key limitations: they require expert demonstrations, which can be difficult or costly to obtain, and they are limited to relatively slow, quasi-static tasks. In this paper, we leverage a tight connection between sampling-based predictive control and generative modeling to address these issues. In particular, we introduce generative predictive control, a supervised learning framework for tasks with fast dynamics that are easy to simulate but difficult to demonstrate. We show how trained flow-matching policies can be warm-started at inference time, maintaining temporal consistency and enabling high-frequency feedback. We believe that generative predictive control offers a complementary approach to existing behavior cloning methods, and hope that it will pave the way toward generalist policies that extend beyond quasi-static demonstration-oriented tasks.