LIBERO

LIBERO runs in the current Python environment through the praxis-libero and praxis-robosuite packages.

Install

pip install "praxis-eval[libero]"

LIBERO config is written under .tmp/libero_config by the evaluator when needed. There is no praxis-eval-setup libero command.

Verify

Run this only on a machine that can import LIBERO and run offscreen MuJoCo:

praxis-eval-verify libero --task libero_10 --task-id 0
praxis-eval-verify libero --help

Task Selection

Default evaluator task: libero_10.

Known suite selectors include:

• libero_spatial
• libero_object
• libero_goal
• libero_10
• comma-separated suites such as libero_spatial,libero_object

Use EvalConfig.task_ids to select task ids inside each suite.

from praxis_eval import EvalConfig

config = EvalConfig(
    task="libero_10",
    task_ids=(0, 1),
    num_eval_per_task=5,
    num_parallel_env=2,
    output_dir="eval/libero_10",
)

Observation Format

Default config values:

Option	Default
obs_type	pixels_agent_pos
camera_name	agentview_image,robot0_eye_in_hand_image
camera_name_mapping	agentview_image -> image, robot0_eye_in_hand_image -> image2
observation_height / observation_width	360 / 360

Policy-facing keys:

Key	Shape / dtype	Notes
task	str	Natural-language LIBERO task instruction.
observation.images.image	(3, H, W), uint8 or float32	Default agent-view RGB camera after preprocessing.
observation.images.image2	(3, H, W), uint8 or float32	Default wrist/in-hand RGB camera after preprocessing.
observation.state.eef_pos	(3,)	Present with obs_type="pixels_agent_pos".
observation.state.eef_quat	(4,)	Present with obs_type="pixels_agent_pos".
observation.state.eef_mat	(3, 3)	Present with obs_type="pixels_agent_pos".
observation.state.gripper_qpos	(2,)	Present with obs_type="pixels_agent_pos".
observation.state.gripper_qvel	(2,)	Present with obs_type="pixels_agent_pos".
observation.state.joint_pos	(7,)	Present with obs_type="pixels_agent_pos".
observation.state.joint_vel	(7,)	Present with obs_type="pixels_agent_pos".

Set obs_type="pixels" to omit the robot-state fields.

Action Format

Field	Value
Shape	(7,)
Dtype	float32
Range	[-1.0, 1.0]
Convention	normalized_delta_pose_gripper

The default control mode is relative. Actions are validated against the 7-D ActionSpec before stepping the environment.

Runtime Notes

• LIBERO uses suite-specific maximum horizons when the suite supplies them.
• The evaluator can use initial states from the installed LIBERO package.
• Rendering is offscreen and depends on the local MuJoCo/OpenGL setup.

Caveats

• LIBERO assets and init states must be available through the installed praxis-libero package.
• Custom cameras require a matching camera_name_mapping.
• The policy adapter is responsible for model-specific image normalization and tokenization.