Licheng Zhong · Gim Hee Lee
| Guide | Contents |
|---|---|
| Installation and data | environment, LIBERO, downloading data and checkpoints, building the data, hardware notes, the clean-git-tree rule for training |
| LeWM-based BC and PPP | world-model pretraining, BC and PPP training, closed-loop evaluation |
| Seer and Seer-PPP | training and closed-loop evaluation |
| Released checkpoints | checkpoint layout, reference results, re-running the evaluation |
Quick start, evaluating the released PPP checkpoint:
conda create -n ppp python=3.10 -y && conda activate ppp
pip install -r requirements.txt --extra-index-url https://download.pytorch.org/whl/cu128
git clone https://github.com/Lifelong-Robot-Learning/LIBERO.git external/LIBERO
git -C external/LIBERO checkout 8f1084e
pip install -e external/LIBERO --no-deps --config-settings editable_mode=compat
python -c "import libero.libero" # answer N once to create ~/.libero/config.yaml
hf download zlicheng/PPP --local-dir . --include "assets/latent_extract/*" "checkpoints/lewm_*"
CELL_TAG=rerun_settle EXTRA_ARGS="--grip_field_eps_steps 1.0 --pose_grip_settle" VIDEO_ARGS="" \
bash runs/close_loop_lewm_libero.sh checkpoints/lewm_ppp_libero "checkpoint_120" 2 "0 1 2 3" 0
python scripts/summarize_official_suites.py --run_dir checkpoints/lewm_ppp_libero \
--cell epseed0_rerun_settle_checkpoint_120This code builds on LeWorldModel, Seer, LIBERO and MAE.
The code is released under the Apache License 2.0. Files adapted from LeWorldModel (MIT) and Seer (Apache-2.0) keep their original licenses; see THIRD_PARTY_NOTICES.md. The MAE weights used by the Seer-based models are under CC BY-NC 4.0.
@article{zhong2026ppp,
title = {Progress Predicts Policy},
author = {Zhong, Licheng and Lee, Gim Hee},
journal = {arXiv preprint arXiv:2611.22491},
year = {2026}
}For more questions, please create an GitHub Issue or contact Licheng Zhong: zlicheng@u.nus.edu
