Action Chunking Extends to Fully Online RL

Master's student Kamil Dybek at the University of Warsaw stated that action chunking in reinforcement learning has mainly been examined in offline or offline-to-online regimes. His post reports that the approach also improves performance in a fully online setting, using Contrastive RL as an example. A side-by-side video attachment shows simulation runs labeled Contrastive RL without Action Chunking next to the version that applies chunking. Assistant Professor Ben Eysenbach at Princeton University reposted the thread. The packet records only the statements made in these public posts and the attached clip description.