Dynamics-Informed Action Learning for Goal-Conditioned Latent Planning

This research paper presents Dynamics-Informed Action Learning (DIAL), a novel approach designed to address the disconnect between learning latent dynamics and utilizing them for planning in visual control tasks. The authors propose training methods that expose forward models to autoregressive predictions and train an inverse dynamics head to infer actions from latent-state transitions. At test time, DIAL couples these dynamics to initialize action sequences for the Cross-Entropy Method (CEM), refining the search towards a goal. Experiments on four visual control benchmarks show that DIAL outperforms existing latent world models and planning baselines in both standard and long-horizon scenarios.