dreamer

Why we included this project

Dreamer is the original implementation of the ICLR 2020 approach that learns control policies by imagining trajectories inside a learned world model. Instead of planning directly in image space, it compresses observations into a compact feature space and predicts ahead there, then backpropagates value gradients through those imagined sequences to train a long-horizon policy. The code is small and self-contained, which makes it a practical reference for reproducing the paper's results on DeepMind Control Suite tasks or adapting the world-model idea to your own simulated control problems. It is a research artifact rather than a production service, so expect to work with TensorFlow 2, dm_control, and the paper instead of a drop-in library. The author points to the newer DreamerV2 repository for Atari and DMControl support, so treat this as the original, simpler implementation.