Generative Models From and For Sampling-Based MPC: A Bootstrapped Approach For Adaptive Contact-Rich Manipulation

2025-10-16

AI Summary

Real-time sampling-based model predictive control (SPC) for quadrupedal robots in contact-rich dexterous manipulation suffers from low sample efficiency, high horizon requirements, and poor task adaptability. Method: This paper proposes a generative predictive control (GPC) framework that directly learns the distribution of control sequences generated by SPC in simulation using a conditional flow matching model, establishing a data-driven generative prior to provide high-quality initial sampling distributions for online planning. GPC requires neither gradient-based optimization nor iterative refinement and enables end-to-end simulation-to-real transfer. Contribution/Results: Experiments demonstrate that GPC significantly improves sample efficiency and planning speed, reduces required prediction horizon, and exhibits strong generalization and robustness across diverse task variants.

Technology Category

• Search and Optimization: Sampling/Simulation-based Search
• Humans and AI: Human-Aware Planning and Behavior Prediction
• Planning, Routing, and Scheduling: Learning for Planning and Scheduling

Application Category

• Economics, Online Markets and Human Computation: Economic ramifications for generative AI infrastructure and applications
• User Modeling, Personalization and Recommendation: User modeling and simulation for interactive and conversational systems
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Abstract

We present a generative predictive control (GPC) framework that amortizes sampling-based Model Predictive Control (SPC) by bootstrapping it with conditional flow-matching models trained on SPC control sequences collected in simulation. Unlike prior work relying on iterative refinement or gradient-based solvers, we show that meaningful proposal distributions can be learned directly from noisy SPC data, enabling more efficient and informed sampling during online planning. We further demonstrate, for the first time, the application of this approach to real-world contact-rich loco-manipulation with a quadruped robot. Extensive experiments in simulation and on hardware show that our method improves sample efficiency, reduces planning horizon requirements, and generalizes robustly across task variations.

Problem

• Amortizing sampling-based MPC with generative models
• Learning proposal distributions from noisy SPC data
• Enabling contact-rich manipulation for quadruped robots

Innovation

• Bootstrapping MPC with conditional flow-matching models
• Learning proposal distributions directly from noisy data
• Applying generative control to real-world contact-rich manipulation