<ul><li>Real-world datasets collected from sensors or human inputs often contain noise and errors, making offline reinforcement learning challenging.</li><li>Existing methods struggle with corruption in high-dimensional state spaces and simultaneous corruption of multiple data elements.</li><li>The proposed Ambient Diffusion-Guided Dataset Recovery (ADG) utilizes diffusion models to address data corruption in offline RL.</li><li>ADG introduces Ambient Denoising Diffusion Probabilistic Models to identify clean and corrupted data, leading to improved offline RL performance.</li></ul>

ADG: Ambient Diffusion-Guided Dataset Recovery for Corruption-Robust Offline Reinforcement Learning

Discover more