Skip to content

INSTALL.md

Latest commit

 

History

History
65 lines (52 loc) · 4.22 KB

INSTALL.md

File metadata and controls

65 lines (52 loc) · 4.22 KB
Raw

Installation

1. Upstream code

StructRL is a set of patches and add-on modules for the upstream projects below, pinned to the commits used in the paper.

Repository Commit Patch Used for
RLinf 4993639 rlinf_core.patch RL training on RoboCasa365
robocasa 8f3c96e robocasa_gymutils.patch RoboCasa365
robosuite 232ce7d (v1.5.2) none RoboCasa365
Isaac-GR00T 4af2b62 isaacgroot_shim.patch GR00T-N1.5
openpi (RoboCasa fork) 5a6beda openpi_rc365.patch π0.5 SFT and evaluation on RoboCasa365
RLinf, second checkout 4993639 rlinf_core.patch, then libero_addon.patch RL training on LIBERO
LIBERO (RLinf fork) 0c5e40c none RL training on LIBERO
LIBERO 8f1084e none LIBERO evaluation client

scripts/setup_upstream.sh clones them next to the code, applies the patches, and installs the Hydra configs into RLinf:

git clone https://github.com/amazon-science/StructRL.git && cd StructRL
export PROJECT_ROOT=$PWD
bash scripts/setup_upstream.sh              # RoboCasa365 with GR00T-N1.5
bash scripts/setup_upstream.sh --openpi     # + π0.5 on RoboCasa365
bash scripts/setup_upstream.sh --libero     # + LIBERO (second RLinf in $PROJECT_ROOT/libero_rl)
export LIBERO_RL_ROOT=$PROJECT_ROOT/libero_rl CKPT_ROOT=$PROJECT_ROOT/ckpts

2. Python environments

The launchers use the virtual environments below, under $PROJECT_ROOT. envs/ has the frozen package list of each one (Python 3.11 unless noted). Install the list, then the patched checkouts it names with pip install --no-deps -e. RLinf itself is used from source through PYTHONPATH.

Environment Package list Used by
.venv-rlinf envs/rlinf-gr00t.txt GR00T-N1.5 training on RoboCasa365
.venv-groot envs/groot-eval.txt GR00T-N1.5 evaluation; policy server for LIBERO evaluation
.venv-rlinf-openpi-train envs/rlinf-openpi-train.txt π0.5 training on RoboCasa365
.venv-rlinf-libero envs/rlinf-libero.txt GR00T-N1.5 training on LIBERO
.venv-rlinf-openpi-libero envs/rlinf-openpi-libero.txt π0.5 training on LIBERO
.venv-libero envs/libero-eval.txt (Python 3.10) LIBERO evaluation client
openpi/.venv uv sync in the openpi fork π0.5 SFT and policy server on RoboCasa365
.venv-rlinf-openpi-eval robocasa, robosuite, openpi-client π0.5 evaluation client on RoboCasa365

3. Data and SFT checkpoints

What Source Location
RoboCasa365 v1.0 datasets RoboCasa docs $PROJECT_ROOT/datasets (see configs/env/)
GR00T-N1.5, RoboCasa365 Composite-Seen SFT robocasa/robocasa365_checkpoints $PROJECT_ROOT/ckpts/gr00t_n1-5/foundation_model_learning/target_posttraining/composite_seen/checkpoint-60000
π0.5, RoboCasa365 SFT train with launch/sft_pi05.sh, then convert to PyTorch export PI05_SFT_CKPT=<dir>
GR00T-N1.5, LIBERO SFT youliangtan/gr00t-n1.5-libero-long-posttrain (and -spatial, -object, -goal) $CKPT_ROOT/gr00t-n1.5-libero-{long,spatial,object,goal}
π0.5, LIBERO SFT RLinf/RLinf-Pi05-LIBERO-130-fullshot-SFT $CKPT_ROOT/RLinf-Pi05-LIBERO-130-fullshot-SFT

LIBERO reads its asset paths from $LIBERO_CONFIG_PATH/config.yaml (default ~/.libero), which LIBERO creates on first import.

4. Cluster

The launchers are Slurm batch scripts: training uses 4 nodes with 8 GPUs each, evaluation 2 nodes with 8 GPUs each. Replace <YOUR_PARTITION>, submit from the repository root (Slurm logs go to logs/), and set NCCL_SOCKET_IFNAME in the NCCL block of each launcher to your network interface prefix.