iwillsolvehardestproblem/dreamerv3-atari100k-breakout

DreamerV3 · Atari-100k Breakout (dream-drift)

To our knowledge the first publicly released DreamerV3 Atari checkpoint. Trained for the dream-drift rollout-drift benchmark.

• Code: official danijar/dreamerv3, unmodified, --configs atari100k --task atari100k_breakout
• Budget: Atari-100k (110k env steps), single NVIDIA L4 (g6.xlarge), ~8.5 h wall clock, jax[cuda12]==0.4.33
• Final training episode scores: 0 to 6 (typical for DreamerV3 Breakout at 100k; see train_Breakout.log)
• Files: ckpt/<snapshot>/agent.pkl (full agent: world model + policy, ~2 GB), ckpt/latest pointer, breakout_dreamer_replay.npz (last ~7k steps of its own replay: obs uint8 (N,64,64,3), act, rew, end, ep_len, num_actions=4), training log.

Load

# inside a danijar/dreamerv3 checkout with its requirements installed
# point --logdir at a folder containing this repo's ckpt/ directory, or see
# harness/dreamer_drift.py in the dream-drift repo for a scoring example.

Measured drift (this checkpoint)

Open-loop 50-step dreams vs. ground truth, 32 starts x 4 samples (protocol in the dream-drift repo): in-distribution k1 MSE 1.2e-3, k50 1.7e-3 (drift 5.3e-4); on expert-policy states 7x worse with narrower sample spread (the "confident healing" failure mode).