GPC Opensources Flow-Matching Robot Policies Trained via Sampling-Based Predictive Control

rsasaki0109 · x · 2026-09-05

Vince Kurtz and Joel Burdick released the code for 'Generative Predictive Control: Flow Matching Policies for Dynamic and Difficult-to-Demonstrate Tasks' (79 GitHub stars). GPC targets robot tasks that are hard to demonstrate but easy to simulate: it alternates between generating training data with sampling-based predictive control, fitting a generative flow-matching model to that data, and using the model to improve the sampling distribution. The repo includes conda setup and training/testing examples across multiple robot systems, offering a reproducible alternative to expensive teleoperated demonstrations.

Original post →

More from Embodied

Embodied channel →