RoboCasa

RoboCasa runs in the current Python environment and requires RoboCasa kitchen assets.

Install

pip install "praxis-eval[robocasa]"

Setup

praxis-eval-setup robocasa
praxis-eval-setup robocasa --dataset-base-path /data/robocasa
praxis-eval-setup robocasa --skip-download
praxis-eval-setup robocasa --help

The setup command downloads kitchen assets unless --skip-download is passed, creates robocasa/macros_private.py when needed, and writes DATASET_BASE_PATH. The default dataset root is $PRAXIS_EVAL_ROBOCASA_DATASET_ROOT or ./data/robocasa.

Verify

Run this only on a machine that can run RoboCasa:

praxis-eval-verify robocasa --task CloseToasterOvenDoor
praxis-eval-verify robocasa --help

Task Selection

Default evaluator task: mt5.

Built-in evaluator group:

Selector	Leaf tasks
mt5	CloseToasterOvenDoor, OpenDrawer, PickPlaceDrawerToCounter, TurnOnElectricKettle, SlideDishwasherRack

You may also pass a RoboCasa365 leaf task such as CloseToasterOvenDoor or a task set advertised by the installed RoboCasa registry.

from praxis_eval import EvalConfig

config = EvalConfig(
    task="mt5",
    num_eval_per_task=5,
    num_parallel_env=2,
    output_dir="eval/robocasa_mt5",
)

Observation Format

Default config values:

Option	Default
image_size	128
camera_names	robot0_agentview_left, robot0_agentview_right, robot0_eye_in_hand
split	all
max_episode_steps	500, overridden by task-specific horizons when known

Policy-facing keys:

Key	Shape / dtype	Notes
task	str	Episode language instruction from RoboCasa metadata.
observation.images.robot0_agentview_left	(3, H, W), usually float32	RGB camera after LeRobot preprocessing.
observation.images.robot0_agentview_right	(3, H, W), usually float32	RGB camera after LeRobot preprocessing.
observation.images.robot0_eye_in_hand	(3, H, W), usually float32	RGB camera after LeRobot preprocessing.
observation.state	(16,), float32	Concatenated official RoboCasa v1.0 state modality.
observation.state.robot0_base_pos	(3,), float32	Per-key alias retained by the processor.
observation.state.robot0_base_quat	(4,), float32	Per-key alias retained by the processor.
observation.state.robot0_base_to_eef_pos	(3,), float32	Per-key alias retained by the processor.
observation.state.robot0_base_to_eef_quat	(4,), float32	Per-key alias retained by the processor.
observation.state.robot0_gripper_qpos	(2,), float32	Per-key alias retained by the processor.

Action Format

Field	Value
Shape	(12,)
Dtype	float32
Range	[-1.0, 1.0] at the public evaluator contract
Convention	robocasa_normalized_mobile_manipulator_action

The evaluator expects the official RoboCasa365 LeRobot action order: base motion, control mode, end-effector position, end-effector rotation, and gripper close. The wrapper maps that order to RoboCasa's native controller order before stepping the simulator and clips to task-specific native controller bounds when needed.

Runtime Notes

• RoboCasa uses fork-based async workers when available.
• Environment construction retries known intermittent MuJoCo layout failures with shifted seeds.
• The wrapper flips RoboCasa camera images to correct robosuite's image orientation.

Caveats

• Asset setup mutates the installed RoboCasa package by writing macros_private.py.
• Large Objaverse visual meshes may be neutralized by a runtime guard to avoid MuJoCo XML failures.
• Some scenes are expensive to build; the eval pool has a long build timeout for retargeting.