mohendy/planet-dmc-rgb-seed0

PlaNet on DeepMind Control Suite (RGB, seed 0)

TensorFlow checkpoints from a one-seed reproduction of Learning Latent Dynamics for Planning from Pixels (PlaNet). Observations are 64 x 64 RGB images. Training used the paper hyperparameters and 1,000 collected episodes per task.

Included checkpoints

Directory	Task	Training checkpoint
cartpole_swingup/	Cartpole Swingup	4,504,500
reacher_easy/	Reacher Easy	4,504,500
cheetah_run/	Cheetah Run	4,504,500

These are the latest checkpoints saved by the training schedule. Training continued to 5M trainer steps, but the final state was not checkpointed.

Post-hoc evaluation

Evaluation uses the paper's CEM settings and ten sequential single-agent trajectories (batch_size=1). This avoids a candidate/state ordering bug in the original planner when multiple agents are planned for in one batch.

Task	Mean	Standard deviation	Paper mean (5 seeds)
Cartpole Swingup	845.16	12.79	821
Reacher Easy	595.40	473.27	832
Cheetah Run	670.43	12.06	662

Reacher has sparse rewards: six of ten trajectories scored 953--1000, while four scored 0--59. These checkpoints use one training seed, whereas the paper reports means over five seeds and ten trajectories per seed.

Exact completed evaluation returns are in posthoc-evaluation.json.

Runtime environment

This reproduction ports the original TensorFlow 1.13 code to TensorFlow 2.15 through tensorflow.compat.v1. It uses dm-control==1.0.9 and mujoco==2.3.2, rather than the paper's 2019 dm_control/MuJoCo 1.50 stack. See environment-reproduction.yml and REPRODUCING.md.

To evaluate from a checkout of the accompanying PlaNet port:

CUDA_VISIBLE_DEVICES='' MUJOCO_GL=egl MUJOCO_EGL_DEVICE_ID=1 \
python planet/scripts/evaluate_checkpoint.py \
  --task cartpole_swingup \
  --run-dir /path/to/cartpole_swingup \
  --episodes 10 --seed 1000

The command above keeps TensorFlow on CPU; the selected GPU is used only for MuJoCo EGL rendering. Full CEM evaluation is substantially faster when the model itself can use an available GPU.

Citation

@inproceedings{hafner2019learning,
  title={Learning Latent Dynamics for Planning from Pixels},
  author={Hafner, Danijar and Lillicrap, Timothy and Fischer, Ian and Villegas,
          Ruben and Ha, David and Lee, Honglak and Davidson, James},
  booktitle={International Conference on Machine Learning},
  year={2019}
}