Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Lara Brudermüller, Brandon Hung, Xinghao Zhu, Jiuguang Wang, Nick Hawes, Preston Culbertson, Simon Le Cleac'h

TL;DR

The paper tackles the challenge of real-time, high-dimensional control for contact-rich loco-manipulation with sampling-based MPC by introducing Generative Predictive Control (GPC). GPC bootstraps online SPC with offline-trained conditional flow-matching models that produce a learned proposal distribution $p_\theta(U|\mathbf{x},\mathbf{h})$, which is used to guide or augment the SPC sampling process. This approach enables non-myopic planning with improved sample efficiency and generalization, demonstrated through both simulated tasks and real hardware on a Spot-based loco-manipulation system. Key findings show that GPC-CEM achieves higher success rates under constrained computational budgets, generalizes to task variations, and reduces planning horizons without sacrificing performance, highlighting the practical impact of integrating learned generative priors into online optimization for robotics.

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.

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

TL;DR

The paper tackles the challenge of real-time, high-dimensional control for contact-rich loco-manipulation with sampling-based MPC by introducing Generative Predictive Control (GPC). GPC bootstraps online SPC with offline-trained conditional flow-matching models that produce a learned proposal distribution , which is used to guide or augment the SPC sampling process. This approach enables non-myopic planning with improved sample efficiency and generalization, demonstrated through both simulated tasks and real hardware on a Spot-based loco-manipulation system. Key findings show that GPC-CEM achieves higher success rates under constrained computational budgets, generalizes to task variations, and reduces planning horizons without sacrificing performance, highlighting the practical impact of integrating learned generative priors into online optimization for robotics.

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.
Paper Structure (17 sections, 4 equations, 4 figures, 2 tables, 1 algorithm)

This paper contains 17 sections, 4 equations, 4 figures, 2 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Background
  4. Problem Formulation
  5. Sampling-based MPC (SPC)
  6. Generative Modeling: Flow-Matching
  7. Bootstrapping Sampling-based MPC with Generative Flow-Matching Models
  8. Data collection
  9. Learning Control Sequence Proposal Distributions
  10. GPC-CEM: Bootstrapping SPC with Flow-Matching
  11. Application to Loco-Manipulation
  12. Experimental Results
  13. Push-T Task
  14. Spot Loco-Manipulation
  15. Limitations
  16. ...and 2 more sections

Figures (4)

  • Figure 1: Generative predictive control (GPC) framework for bootstrapping sampling-based Model Predictive Control (SPC). We collect open-loop control sequences from an SPC algorithm in simulation and use them to train a generative proposal distribution. At test time, this model guides and amortizes online MPC, enabling non-myopic, constraint-satisfying behavior with improved sample efficiency and robustness in contact-rich, high-dimensional settings.
  • Figure 2: The mapping from the high-level action space to the low-level spot loco-manipulation control space. The locomotion policy is pre-trained to follow the high-level commands while maintaining balance and stability. It is fixed throughout the full task planning and execution.
  • Figure 3: Push-T Task overview.Left: Original Push-T task chi2023diffusion, where a circular robot is required to push a T-shaped block into a target pose, shown in green. Right: Modified task with a K- instead of T-block at the bottom.
  • Figure 5: Spot Loco-Manipulation Experiment: Qualitative examples of two successful Spot loco-manipulation task runs in the real world with GPC-CEM. Both images show a overlay of several snapshots of the trajectories of the robot and the chair.