DR-APO: Decoupled-Representation Advantage-Guided Policy Optimization from Continuous Control to Embodied Navigation

This research proposes DR-APO, a policy optimization method that separates policy improvement from representation, allowing for unified optimization across Gaussian, diffusion, and flow policies. Evaluated on MuJoCo-v5 tasks, DR-APO achieves competitive results with SAC for Gaussian policies and sets new benchmarks for diffusion and flow policies. In embodied navigation, NavDP post-training improves success rates from 70.33% to 89.55% in simulation and from 25% to 65% in real-world settings.