Update, Nov 20 2024: We fixed the issue of EMA update being too infrequent causing slow pre-training (commit). Now the number of epochs needed for pre-training can be much lower than those used in the configs (e.g., 3000 for robomimic state and 1000 for robomimic pixel), and we have updated the pre-training configs in v0.7. If you would like to replicate the original experimental results from the paper, please use v0.6.
Gym configs are under cfg/gym/pretrain/<env_name>/, and the config name is pre_diffusion_mlp. Robomimic configs are under cfg/robomimic/pretrain/<env_name>/, and the name is also pre_diffusion_mlp.
Note: In both Gym and Robomimic experiments, for the experiments in the paper we used more than enough expert demonstrations for pre-training. You can specify +train_dataset.max_n_episodes=<number_of_episodes> to limit the number of episodes so the pre-training is faster.
Robomimic configs are under cfg/robomimic/pretrain/<env_name>/, and the naming follows pre_<diffusion/gaussian/gmm>_<mlp/unet/transformer>_<img?>. Furniture-Bench configs are under cfg/furniture/pretrain/<env_name>/, and the naming follows pre_<diffusion/gaussian>_<mlp/unet>.
D3IL configs are under cfg/d3il/pretrain/avoid_<mode>/, and the naming follows pre_<diffusion/gaussian/gmm>_mlp. In the paper we manually examine the pre-trained checkpoints and pick the ones that visually match the pre-training data the best. We also tune the Gaussian and GMM policy architecture extensively for best pre-training performance. The action chunk size can be specified with horizon_steps and the number of denoising steps can be specified with denoising_steps.