Installation

Evaluation shares the same embodied environment installation flow as training. From the repository root:

bash requirements/install.sh embodied --model <model> --env <env>
source .venv/bin/activate

Choose <model> and <env> to match your target benchmark:

Benchmark	Recommended --model	Recommended --env
LIBERO	openpi / openvla-oft / starvla / dreamzero	maniskill_libero or libero
RoboTwin	openvla-oft / openpi / lingbotvla	robotwin
BEHAVIOR-1K	openpi	behavior
ManiSkill OOD	openvla-oft	maniskill_libero
RealWorld	openpi / dreamzero	franka
PolaRiS	openpi	polaris

For more installation options, see Installation.

Benchmark-Specific Environment Variables

Configure the following when your benchmark requires them (LIBERO is handled automatically by install.sh; see below):

LIBERO

With --env libero, install.sh clones LIBERO into .venv/libero by default and appends it to PYTHONPATH in the activate script. To use an existing checkout instead, set LIBERO_PATH before installation:

export LIBERO_PATH=/path/to/LIBERO

RoboTwin

export ROBOTWIN_PATH=/path/to/RoboTwin
export ROBOT_PLATFORM=ALOHA

BEHAVIOR-1K

Set OMNIGIBSON_DATA_PATH and related OmniGibson paths. See RL with Behavior Benchmark.

DreamZero

export DREAMZERO_PATH=/path/to/DreamZero

PolaRiS

export POLARIS_DATA_PATH=/path/to/dataset/PolaRiS-Hub

run_eval.sh reads these variables automatically and configures PYTHONPATH.

Next Steps

After installation, follow Quick Tour for your first evaluation, or see the benchmark guide: Benchmark Guides.