Reparameterization Flow Policy Optimization
This work addresses the problem of improving sample efficiency and performance in model-based reinforcement learning for robotics tasks, representing an incremental advancement by combining flow policies with existing reparameterization gradients.
The paper tackled the limitation of prior reparameterization policy gradient methods to Gaussian policies by proposing Reparameterization Flow Policy Optimization (RFO), which integrates flow policies with differentiable dynamics for model-based reinforcement learning, achieving almost 2x the reward on a challenging soft-body quadruped locomotion task compared to the state-of-the-art baseline.
Reparameterization Policy Gradient (RPG) has emerged as a powerful paradigm for model-based reinforcement learning, enabling high sample efficiency by backpropagating gradients through differentiable dynamics. However, prior RPG approaches have been predominantly restricted to Gaussian policies, limiting their performance and failing to leverage recent advances in generative models. In this work, we identify that flow policies, which generate actions via differentiable ODE integration, naturally align with the RPG framework, a connection not established in prior work. However, naively exploiting this synergy proves ineffective, often suffering from training instability and a lack of exploration. We propose Reparameterization Flow Policy Optimization (RFO). RFO computes policy gradients by backpropagating jointly through the flow generation process and system dynamics, unlocking high sample efficiency without requiring intractable log-likelihood calculations. RFO includes two tailored regularization terms for stability and exploration. We also propose a variant of RFO with action chunking. Extensive experiments on diverse locomotion and manipulation tasks, involving both rigid and soft bodies with state or visual inputs, demonstrate the effectiveness of RFO. Notably, on a challenging locomotion task controlling a soft-body quadruped, RFO achieves almost $2\times$ the reward of the state-of-the-art baseline.