Improving Diffusion Planners by Self-Supervised Action Gating with Energies
Researchers have developed a new method called Self-supervised Action Gating with Energies (SAGE) to improve diffusion planners used in offline reinforcement learning. SAGE works by penalizing plans that are inconsistent with environmental dynamics, using a latent consistency signal derived from a Joint-Embedding Predictive Architecture (JEPA) encoder. This approach integrates into existing diffusion planning pipelines without requiring environment rollouts or policy retraining, enhancing performance and robustness across various benchmarks. AI
IMPACT Enhances reinforcement learning planning by improving robustness and performance without additional training.