Configs

We adopt hydra and provide a configuration file configs/default.yaml to configure the settings before running experiments. Here we explain some notable terms.

• Path

  • base_root. Base directory.

  • data_root. The directory of data.

  • asset_root. The directory of asset samples and materials.

  • output_root. The directory to save experiment outputs.

• Task and model

  • mode. Either train or eval.

  • task. Task name for single-task, multi for multi-task.

  • model. One of cliport6d, peract, bc_lang_cnn, bc_lang_vit.

  • lang_encoder. One of clip, none, t5, roberta.

  • state_head. Whether using an additional state head, either 0 or 1.

• Running arguments

  • batch_size. 8 for training and 1 for evaluation.

  • steps. Training steps, default at 100k.

  • log_interval. The step interval between logging behavior during training.

  • save_interval. The step interval between checkpoint saving during training.

  • use_gt. For evaluation, two bool values indicating whether to use ground-truth keypoints for each phase.

  • visualize: For evaluation, keep rendering if True. Setting to False can accelerate evaluation.

• Environment

  • offset_bound. The perception bound, used to crop a cube centered at the robot, represented in [x1, y1, z1, x2, y2, z2], in m.

  • iso_surface. Whether enabling realistic fluid simulation.

• PerAct

  • t5_cfg. The path of pre-trained T5 model.

  • roberta_cfg. The path of pre-trained RoBERTa model.

In addition to configuring the settings in yaml file, we can also configure them when running commands, e.g.,

python eval.py task=pickup_object model=peract lang_encoder=clip mode=eval use_gt=[0,0] visualize=0