Supported Models

Eval configs reference model presets from examples/embodiment/config/model/ via defaults, and override fields such as model_path under rollout.model. Models with example configs in evaluations/ today:

Model	model_type	Example config
π₀ / π₀.₅ (OpenPI)	openpi	libero_spatial_openpi_pi05_eval, libero_goal_openpi_eval, robotwin_adjust_bottle_openpi_eval, etc.
OpenVLA-OFT	openvla_oft	libero_10_openvlaoft_eval, robotwin_place_empty_cup_openvlaoft_eval, maniskill_ood_openvlaoft_eval, etc.
StarVLA	starvla	libero_spatial_starvla_eval
DreamZero	dreamzero	libero_spatial_dreamzero_eval, realworld_pnp_eval_dreamzero
LingBotVLA	lingbotvla	robotwin_click_bell_lingbotvla_eval, robotwin_place_shoe_lingbotvla_eval

Model Path

Set rollout.model.model_path in YAML or on the command line to a local model directory. For .pt checkpoints from RL training, also set runner.ckpt_path.

Standalone Eval Scripts

Some models also provide standalone eval scripts under toolkits/standalone_eval_scripts/. These follow specific evaluation protocols but do not support parallel environment acceleration. Prefer evaluations/run_eval.sh when possible.