
Dreamer
Latent statecartpole swing-up · DeepMind Control
Return 75 → 493
60k environment steps · 4 GB laptop GPU
Every model here was trained from scratch on one 4 GB laptop GPU, then recorded: the real world on the left, the model’s imagination on the right.

cartpole swing-up · DeepMind Control
Return 75 → 493
60k environment steps · 4 GB laptop GPU

cartpole swing-up · DeepMind Control
Return 75 → 165–250
60k environment steps · 4 GB laptop GPU

cartpole swing-up · DeepMind Control
Return 76 → 435
75 minutes · 4 GB laptop GPU

cartpole swing-up, random actions
GRU and LSTM backbones side by side
25 minutes per backbone · 4 GB laptop GPU

Breakout · Atari
Draws everything but the ball
21 epochs · 4 GB laptop GPU

Boxing · Atari
Plays a full round from 16 tokens a frame
300 epochs · 4 GB laptop GPU

Sonic · TinyWorlds video
Learns actions from unlabelled video
6k steps · 4 GB laptop GPU

CIFAR-10 images
Draws images class by class
40 epochs · 4 GB laptop GPU

CIFAR-10 images
41.1% kNN accuracy vs 29.6% untrained
30 epochs · 4 GB laptop GPU
python demos/dreamer_demo.py trainpython demos/dreamer_demo.py recordpython demos/run_all.pyIRIS is the exception: its full schedule runs for several days on 4 GB.
Open source, MIT licensed, and runs on a laptop GPU.
pip install simulusion