[Paper] [Project Website]

In this work, we propose a novel view that treats inducing temporal action abstractions as a sequence compression problem. To do so, we bring a subtle but critical component of LLM training pipelines -- input tokenization via byte pair encoding (BPE) -- to the seemingly distant task of learning skills of variable time span in continuous control domains.

We introduce an approach called Primitive Sequence Encoding (PRISE) that combines continuous action quantization with BPE to learn powerful action abstractions. We empirically show that high-level skills discovered by PRISE from a multitask set of robotic manipulation demonstrations significantly boost the performance of both multitask imitation learning as well as few-shot imitation learning on unseen tasks.

This thread contains implementation of PRISE on LIBERO. Check out the MetaWorld thread to see PRISE implementation on MetaWorld with slight modification compared to LIBERO.

Step 1: Setup Environment:

conda create -n libero python=3.8.13
conda activate libero
pip install -r requirements.txt
pip install -e .

Step 2: Install LIBERO:

git clone https://github.com/Lifelong-Robot-Learning/LIBERO.git
cd LIBERO
pip install -r requirements.txt
pip install torch==1.11.0+cu113 torchvision==0.12.0+cu113 torchaudio==0.11.0 --extra-index-url https://download.pytorch.org/whl/cu113
pip install -e .

Step 3: Download and Preprocess LIBERO Dataset: First, download the LIBERO dataset.

cd LIBERO
python benchmark_scripts/download_libero_datasets.py --datasets libero_100

Next, we need to convert the format of the LIBERO dataset. Copy the convert_data.py file into the LIBERO repo, and then run

python convert_data.py --save_path ${DATASET_PATH} --libero_path ${LIBERO_PATH}/LIBERO/libero/datasets/libero_90
python convert_data.py --save_path ${DATASET_PATH} --libero_path ${LIBERO_PATH}/LIBERO/libero/datasets/libero_10

Here ${DATASET_PATH} is the path of the libero dataset that you are going to store. After you converting the data, rewrite libero_path in cfgs/prise_config.yaml to ${LIBERO_PATH}, and data_storage_dir to ${DATASET_PATH}.

Stage I: Pretrain PRISE action vector quantization:

python train_prise.py exp_name=${EXP_NAME} stage=1 n_code=${N_CODE} save_snapshot=true &

The model checkpoint and loss information is saved under the directory exp_local/${EXP_NAME}. By default, we set number of quantized codes to be 10. In terms of the computational resources to train the first stage of PRISE, we use 4 NVIDIA A100 GPU w. 40G memory and 400 GB CPU memory.

Stage II: Run BPE tokenization algorithm to get skill tokens:

python train_prise.py exp_name=${EXP_NAME} stage=2 vocab_size=${VOCAB_SIZE} &

By default, we set voabulary size to be 200. The learned BPE tokenizer will be saved under the directory exp_local/${EXP_NAME}/vocab_libero90_code${N_CODE}_vocab${VOCAB_SIZE}_minfreq10_maxtoken20.pkl. The second stage should take ~10 minutes to finish on a single A100 GPU.

Stage III: Downstream Adaptation:

Case I: Multitask Learning: To train a multitask generalist policy on LIBERO-90:

python train_prise.py exp_name=${EXP_NAME} replay_buffer_num_workers=4 stage=3 exp_bc_name=${DOWNSTREAM_EXP_NAME} multitask=true downstream_task_suite=libero_90 num_train_steps=30010 eval=false save_snapshot=true vocab_size=${VOCAB_SIZE} &

To evaluate the trained multitask policy:

python eval_libero90.py exp_name=${EXP_NAME} &

Case II: Few-shot Adaptation to unseen tasks (5-shots): To adapt to an unseen task with five expert demonstration trajectories:

python train_prise.py exp_name=${EXP_NAME} replay_buffer_num_workers=4 batch_size=64 stage=3 exp_bc_name=${DOWNSTREAM_EXP_NAME} downstream_task_name=${TASK_ID} downstream_task_suite=libero_10 num_train_steps=30010 eval_freq=2000 max_traj_per_task=5 vocab_size=${VOCAB_SIZE} &

If you find our method or code relevant to your research, please consider citing the paper as follows:

@misc{zheng2024prise,
      title={PRISE: Learning Temporal Action Abstractions as a Sequence Compression Problem}, 
      author={Ruijie Zheng and Ching-An Cheng and Hal Daumé III au2 and Furong Huang and Andrey Kolobov},
      year={2024},
      eprint={2402.10450},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}