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` / `molmoact2`	`maniskill_libero` or `libero`
LIBERO / LIBERO-Plus	`fastwam`	`libero` or `liberoplus`
RoboTwin	`openvla-oft` / `openpi` / `lingbotvla` / `lingbotvlav2`	`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:

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.

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