Version 4 was posted on October 21, 2026; the paper was first submitted on March 18.
FMER uses ODE-based flow matching for reinforcement-learning policies. The authors derive an entropy objective that accounts for the density change introduced by tanh when bounded actions are produced from an unbounded latent space. They also state that minimizing an advantage-weighted conditional flow-matching loss is a simulation-free surrogate for policy mirror descent, avoiding backpropagation through the ODE solver. The method tunes its mirror-descent temperature from effective sample size to maintain a trust region.
The paper reports superior results on sparse-reward FrankaKitchen and competitive results on dense-reward MuJoCo benchmarks. It says code is available. The source does not establish performance beyond those reported environments, independent replication, or readiness for production RL systems.