Following DOWNLOAD_DATASET and DOWNLOAD_MODEL, we have prepared the required datasets and models. Then, before training, we need to modify:
- The directory of data:
root_dir
- The path of the language embedding model:
embedding_model_path
These configurations can be updated in libero_exp/configs/base/data/default.yaml.
Please refer to scipts files of LIBERO. To start an experiment, please choose:
Benchmark
from['libero_spatial', 'libero_object', 'libero_goal', 'libero_10']
Policy
from['bc_policy', 'bc_ib_policy']
Backbone
from['mlp', 'rnn', 'transformer', 'vilt']
# for bc_policy
# benchmark policy backbone train_ratio seed
bash libero_exp/scripts/main_libero.sh 'libero_spatial' 'bc_policy' 'transformer' 0.9 0
bash libero_exp/scripts/main_libero.sh 'libero_object' 'bc_policy' 'vilt' 0.9 0
bash libero_exp/scripts/main_libero.sh 'libero_goal' 'bc_policy' 'rnn' 0.9 0
bash libero_exp/scripts/main_libero.sh 'libero_10' 'bc_policy' 'mlp' 0.9 0
# for bc_ib_policy
bash libero_exp/scripts/main_libero.sh 'libero_spatial' 'bc_ib_policy' 'transformer' 0.9 0
bash libero_exp/scripts/main_libero.sh 'libero_object' 'bc_ib_policy' 'vilt' 0.9 0
bash libero_exp/scripts/main_libero.sh 'libero_goal' 'bc_ib_policy' 'rnn' 0.9 0
bash libero_exp/scripts/main_libero.sh 'libero_10' 'bc_ib_policy' 'mlp' 0.9 0
If using eval_libero.sh, the command is as follows:
# diectory of checkpoint only evalute on final checkpoint
bash libero_exp/scripts/eval_libero.sh 'outputs/libero/bc_policy/vilt/libero_goal/1130_1137_seed0' False
If using eval_libero_all.sh, you need to specify the directory of the testing model in the evaluation script before running the evaluation:
# only evalute on final checkpoint
bash libero_exp/scripts/eval_libero_all.sh' False
Following DOWNLOAD_DATASET and DOWNLOAD_MODEL, we have prepared the required datasets and models. Then, before training, we need to modify:
- The directory of data:
data_dir
. It can be updated in data config for Adroit, DMControl, MetaWorld, and Trifinger. - The directory of all pre-trained image embedding model:
embedding_dir
. It can be updated in policy config for Adroit (spatial_fuse and temporal_fuse), DMControl (spatial_fuse and temporal_fuse), MetaWorld (spatial_fuse and temporal_fuse), and Trifinger (spatial_fuse and temporal_fuse).
Please refer to scipts files of CortexBench. To start an experiment, please choose:
Policy
from['bc_policy', 'bc_ib_policy']
Backbone
from['ResNet', 'ViT', 'R3M', 'MVP', 'VC1', 'Voltron', 'MPI']
Fusion Method
from['spatial_fuse', 'temporal_fuse']
Fine-tuning Method
from['full_ft', 'partial_ft']
# for bc_policy
# benchmark policy fine-tuning and fusion backbone seed
bash cortexbench_exp/scripts/main_cortex_metaworld.sh bc_policy full_ft_temporal_fuse ResNet 0
bash cortexbench_exp/scripts/main_cortex_metaworld.sh bc_policy full_ft_spatial_fuse ResNet 0
bash cortexbench_exp/scripts/main_cortex_metaworld.sh bc_policy partial_ft_temporal_fuse VC1 0
bash cortexbench_exp/scripts/main_cortex_metaworld.sh bc_policy partial_ft_spatial_fuse VC1 0
# for bc_ib_policy
bash cortexbench_exp/scripts/main_cortex_metaworld.sh bc_ib_policy full_ft_temporal_fuse ResNet 0
bash cortexbench_exp/scripts/main_cortex_metaworld.sh bc_ib_policy full_ft_spatial_fuse ResNet 0
bash cortexbench_exp/scripts/main_cortex_metaworld.sh bc_ib_policy partial_ft_temporal_fuse VC1 0
bash cortexbench_exp/scripts/main_cortex_metaworld.sh bc_ib_policy partial_ft_spatial_fuse VC1 0
Before evaluation, you need to specify the directory of the testing model in the evaluation script, such as eval_cortex_metaworld.sh
bash cortexbench_exp/scripts/eval_cortex_metaworld.sh