RoboMimic

RoboMimic runs robosuite tasks in the current Python environment.

Install

pip install "praxis-eval[robomimic]"

There is no praxis-eval-setup robomimic command.

Verify

Run this only on a machine that can run robosuite:

praxis-eval-verify robomimic --task Lift --disable-render
praxis-eval-verify robomimic --help

Remove --disable-render when verifying offscreen rendering:

praxis-eval-verify robomimic --task Lift --video-resolution 128

Task Selection

Default evaluator task: mt3.

Selector	Leaf tasks
mt3	Lift, PickPlaceCan, NutAssemblySquare
mt4	Lift, PickPlaceCan, ToolHang, NutAssemblySquare

Leaf tasks and aliases are also supported:

• Lift or lift
• PickPlaceCan, can, or pick_place_can
• NutAssemblySquare, square, or nut_assembly_square
• ToolHang or tool_hang

Observation Format

Default config values:

Option	Default
image_size	128
camera_names	agentview, robot0_eye_in_hand
state_ports	robot0_eef_pos, robot0_eef_quat, robot0_gripper_qpos
video_camera	agentview
max_episode_steps	800, overridden by known task horizons
robot	Panda

Policy-facing keys:

Key	Shape / dtype	Notes
task	str	Natural-language instruction such as Lift the cube.
observation.images.agentview	(3, H, W), usually float32	RGB camera after preprocessing.
observation.images.robot0_eye_in_hand	(3, H, W), usually float32	RGB camera after preprocessing.
observation.state	(9,), float32	Concatenated default state ports.
observation.state.robot0_eef_pos	(3,), float32	Per-port alias retained by the processor.
observation.state.robot0_eef_quat	(4,), float32	Per-port alias retained by the processor.
observation.state.robot0_gripper_qpos	(2,), float32	Per-port alias retained by the processor.

If you customize state_ports, the flat state dimension changes. Known state ports include robot0_joint_pos, robot0_eef_pos, robot0_eef_quat, and robot0_gripper_qpos.

Action Format

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

The wrapper validates shape, finite values, and action-space bounds before calling robosuite.

Runtime Notes

• MUJOCO_GL=egl is set by the verifier when rendering is enabled.
• The wrapper uses the RoboMimic v1.5 PH-style robosuite controller setup.
• Known horizons are Lift=100, PickPlaceCan=200, NutAssemblySquare=300, and ToolHang=800.

Caveats

• Rendering requires a working offscreen MuJoCo/OpenGL setup.
• The default state contract is narrower than all robosuite raw observations; customize state_ports if your policy needs additional proprioception.