Improving Diffusion Planners by Self-Supervised Action Gating with Energies

Researchers propose SAGE (Self-supervised Action Gating with Energies), a new method to improve diffusion planners in offline reinforcement learning by filtering out dynamically inconsistent trajectories. The approach uses a latent consistency signal to re-rank candidate actions at inference time, improving performance across locomotion, navigation, and manipulation tasks without requiring environment rollouts or policy retraining.

• →SAGE addresses brittleness in diffusion planners by penalizing dynamically inconsistent plans using latent prediction errors.
• →The method integrates into existing diffusion planning pipelines without requiring environment rollouts or policy retraining.
• →SAGE uses a Joint-Embedding Predictive Architecture (JEPA) encoder trained on offline state sequences for consistency evaluation.
• →Performance improvements were demonstrated across locomotion, navigation, and manipulation benchmarks.
• →The approach combines feasibility scores with value estimates to make better action selections at test time.