PlaNet

PlaNet (Deep Planning Network) is a model-based reinforcement learning agent developed by researchers at Google and DeepMind (Danijar Hafner et al.), introduced in 2019 in the paper “Learning Latent Dynamics for Planning from Pixels” (ICML 2019). The agent learns a world model from images (pixels) only and uses it to plan actions.

At its core is a Recurrent State-Space Model (RSSM) — a latent dynamics model combining deterministic and stochastic components, letting it predict multiple possible futures and retain information over many time steps. Planning happens in latent space: the agent compares thousands of imagined action sequences, picks the best, executes it, and replans after every new observation.

PlaNet was trained on continuous-control tasks from the DeepMind Control Suite (cartpole swingup, finger spin, cheetah run, cup catch, walker walk, reacher). It reaches performance close to the model-free D4PG while using roughly 50x fewer environment interactions. Reference code was released under the Apache 2.0 license; PlaNet is a predecessor of the Dreamer line of models.