Dynamic Mixture of Progressive Parameter-Efficient Expert Library for Lifelong Robot Learning

Yuheng Lei, Sitong Mao, Shunbo Zhou, Hongyuan Zhang, Xuelong Li, Ping Luo

[Paper] [Pretraining Checkpoint (LIBERO-90)]

Installation

Please run the following commands in the given order to install the dependency and the LIBERO benchmark.

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

Then install the libero package:

pip install -e .

We leverage high-quality human teleoperation demonstrations for the task suites in LIBERO. To download the demonstration dataset, run:

python benchmark_scripts/download_libero_datasets.py

For a detailed walk-through of the LIBERO benchmark, please either refer to the documentation or the original paper.

Training

We can starting training by running:

export CUDA_VISIBLE_DEVICES=GPU_ID && \
export MUJOCO_EGL_DEVICE_ID=GPU_ID && \
python libero/lifelong/main.py seed=SEED \
                               benchmark_name=BENCHMARK \
                               policy=POLICY \
                               lifelong=ALGO

Pretraining

• BENCHMARK from [LIBERO_90]
• ALGO from [multitask]
• POLICY from [bc_foundation_policy_fft, bc_foundation_policy_frozen]

We provide the template script of pretraining as follows:

sh exp_scripts/pretraining_scripts/run_chunkonlyfft_base_clip.sh

Lifelong Learning

• BENCHMARK from [LIBERO_SPATIAL, LIBERO_OBJECT, LIBERO_GOAL, LIBERO_10]
• ALGO from [base, er, ewc, packnet, lotus, l2m, iscil, tail, dmpel]
• POLICY from [bc_foundation_policy_fft, bc_foundation_policy_frozen, bc_hierarchical_policy, bc_foundation_tail_policy, bc_foundation_l2m_policy, bc_foundation_iscil_policy, bc_foundation_dmpel_policy]

We provide the scripts to reproduce results in the paper in exp_scripts/lifelong_scripts. For example, we can evaluate DMPEL in LIBERO-Goal by running:

sh exp_scripts/lifelong_scripts/dmpel.sh

Note that the pretrained model path should be the same as the final checkpoint you saved during pretraining. We also provide our pretraining checkpoint to facilitate the replication of results presented in the main paper.

Acknowledgements

This codebase is built with reference to the following repositories:

• LIBERO
• L2M
• LOTUS
• IsCiL