MetaWorld

MetaWorld runs in the current Python environment through LeRobot's MetaWorld dependency set.

Install

pip install "praxis-eval[metaworld]"

There is no praxis-eval-setup metaworld command.

Verify

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

praxis-eval-verify metaworld --task reach-v3
praxis-eval-verify metaworld --help

Task Selection

Default evaluator task: mt50.

Selectors:

• mt50 or mt-50
• difficulty groups: easy, medium, hard, very_hard
• a leaf task name from LeRobot's metaworld_config.json, such as reach-v3
• comma-separated selectors such as easy,medium

from praxis_eval import EvalConfig

config = EvalConfig(
    task="reach-v3",
    num_eval_per_task=10,
    output_dir="eval/metaworld_reach",
)

Observation Format

Default config values:

Option	Default
obs_type	pixels_agent_pos
camera_name	corner2
observation_height / observation_width	480 / 480
episode_length	500

Policy-facing keys:

Key	Shape / dtype	Notes
task	str	Task instruction from LeRobot MetaWorld metadata.
observation.images.corner2	(3, H, W), uint8 or float32	Default image key. Custom camera_name values use observation.images.<camera_name>.
observation.state	(4,), float32 or float64	First 4 elements of the raw MetaWorld observation when obs_type="pixels_agent_pos".

Set obs_type="pixels" to omit agent position state.

Action Format

Field	Value
Shape	(4,)
Dtype	float32
Range	[-1.0, 1.0]
Convention	metaworld_xyz_gripper

Actions are validated against the normalized MetaWorld action space before stepping.

Runtime Notes

• The wrapper lazily constructs the backend MetaWorld environment on first reset or render.
• The default corner2 camera position is adjusted to match the LeRobot dataset convention.
• The wrapper flips corner2 rendered images to match expected orientation.

Caveats

• The exact list of leaf tasks comes from the installed LeRobot MetaWorld metadata.
• Rendering requires a working MuJoCo/OpenGL setup.