Continual Learning Bash Scripts

This directory contains utility scripts for generating training and evaluation commands, as well as Slurm submission scripts for cluster environments.

See README_GR00T.md in the root directory for more information on GR00T installation and setup.

Configuration

Before running any scripts, ensure that your paths are correctly configured in bash_scripts/config.sh. This file is sourced by all other scripts to ensure consistent dataset and output locations.

# bash_scripts/config.sh
export LIBERO_DATASET_PATH="/path/to/libero/dataset"
export OUTPUT_PATH="/path/to/experiments/output"

---

Command Generation Scripts

These scripts are used to generate the full Python commands required to run experiments. You can pipe the output to a file or execute it directly.

1. Training Command Generation

generate_train_commands.sh

Usage:

./bash_scripts/generate_train_commands.sh <TASK_SUITE_NAME> <STEPS> [CL_TYPE] [N_MEMORIES] [SAVE_STEPS] [SEED] [TUNE_LLM] [TUNE_VISION] [ADD_NAME]

• Example: ./bash_scripts/generate_train_commands.sh libero_10 10000 ER 1000 10000 42 False False s42
• Arguments:
  • TASK_SUITE_NAME: libero_10, libero_object, libero_spatial, libero_goal, libero_90
  • STEPS: Total training steps.
  • CL_TYPE: (Optional) Continual Learning type (default: ER).
  • N_MEMORIES: (Optional) Memory buffer size for ER (default: 1000).

2. Evaluation Command Generation

generate_eval_commands.sh

Usage:

./bash_scripts/generate_eval_commands.sh <TASK_SUITE_NAME> <STEPS> [CL_TYPE] [N_MEMORIES] [SAVE_STEPS] [SEED] [TUNE_LLM] [TUNE_VISION] [ADD_NAME] [CURRENT_TASK_ID] [CKPT_STEP]

• Example: ./bash_scripts/generate_eval_commands.sh libero_10 10000 ER 1000 10000 42 False False s42 6 10000
• Generates evaluation commands for all tasks up to CURRENT_TASK_ID using the checkpoint at CKPT_STEP.
• For the above command, it will evaluate task 6 checkpoint using step 10000, on all tasks from 0 to 6.

---

Slurm Submission Scripts

For running on a cluster using the Slurm workload manager.

1. Slurm Training

slurm_train.sh

Usage:

sbatch bash_scripts/slurm_train.sh <TASK_SUITE_NAME> <STEPS> <CL_TYPE> [N_MEMORIES] [SAVE_STEPS] [SEED] [TUNE_LLM] [TUNE_VISION] [ADD_NAME]

2. Slurm Evaluation

slurm_eval.sh

Uses Slurm job arrays to parallelize evaluation across tasks.

Usage:

sbatch bash_scripts/slurm_eval.sh <TASK_SUITE_NAME> <CL_TYPE> <CKPT_STEP> [N_MEMORIES] [BATCH_SIZE] [ADD_NAME]

---

Note on Python Environment

When running these scripts, ensure your PYTHONPATH includes the project root:

export PYTHONPATH=$PYTHONPATH:$(pwd)

The scripts automatically attempt to set this when pulling configuration from gr00t/config.py.