Quick Tour#
This tutorial walks through your first embodied evaluation in about 5 minutes using LIBERO Spatial + OpenPI π₀.₅. The example config is evaluations/libero/libero_spatial_openpi_pi05_eval.yaml.
Step 1: Install the Environment#
From the repository root:
bash requirements/install.sh embodied --model openpi --env libero
source .venv/bin/activate
With --env libero, the installer clones LIBERO into .venv/libero (or reuses an existing checkout when LIBERO_PATH is set) and appends it to PYTHONPATH in .venv/bin/activate. No separate LIBERO_PATH setup is required.
Step 2: Prepare the Model#
Download the SFT checkpoint from Hugging Face (RLinf/RLinf-Pi05-LIBERO-SFT) to a local directory:
huggingface-cli download RLinf/RLinf-Pi05-LIBERO-SFT --local-dir ./RLinf-Pi05-LIBERO-SFT
Model hub: RLinf/RLinf-Pi05-LIBERO-SFT. You can also override rollout.model.model_path on the command line at launch time.
Step 3: Launch Evaluation#
bash evaluations/run_eval.sh libero libero_spatial_openpi_pi05_eval \
rollout.model.model_path=./RLinf-Pi05-LIBERO-SFT
When the config name starts with libero_, you can omit the benchmark argument:
bash evaluations/run_eval.sh libero_spatial_openpi_pi05_eval \
rollout.model.model_path=./RLinf-Pi05-LIBERO-SFT
Step 4: Check Results#
The terminal prints metrics such as
eval/success_onceandeval/returnLog directory:
logs/<timestamp>-libero_spatial_openpi_pi05_eval/eval_embodiment.logWhen
env.eval.video_cfg.save_video: True, videos are saved under<log_path>/video/eval/
See Logs and Results for more details.
Next Steps#
YAML configuration: Configuration Reference
Other benchmarks: LIBERO Evaluation
More CLI options: CLI Reference