MS-HAB

MS-HAB is evaluated through a dedicated runtime. The caller's policy environment does not need to import the simulator stack.

Install

pip install "praxis-eval[mshab]"

This extra installs praxis-mshab and praxis-remote in the caller environment. The simulator itself usually runs in a dedicated conda or micromamba environment created by setup.

Setup

praxis-eval-setup mshab
praxis-eval-setup mshab --env-manager micromamba --env-name mshab-praxis
praxis-eval-setup mshab --ms-asset-dir /data/mshab/maniskill_assets
praxis-eval-setup mshab --skip-assets
praxis-eval-setup mshab --help

Setup reads the env spec from the installed praxis-mshab package resource mshab/praxis_conda_env.yaml, installs praxis-eval, praxis-remote, and praxis-mshab into that runtime, downloads ManiSkill assets, and writes Vulkan/EGL activation hooks.

Verify

Run this only on a machine with the dedicated MS-HAB runtime:

praxis-eval-verify mshab --env-name mshab-praxis --num-episodes 1 --num-envs 1
praxis-eval-verify mshab --env-python-bin /path/to/mshab-praxis/bin/python
praxis-eval-verify mshab --help

Task Selection

Default evaluator task: set_table.

Selector	Subtask	Target	Task id
pick or set_table_pick	pick	all	0
place or set_table_place	place	all	1
open_fridge or set_table_open_fridge	open	fridge	2
open_kitchen_counter or set_table_open_kitchen_counter	open	kitchen_counter	3
close_fridge or set_table_close_fridge	close	fridge	4
close_kitchen_counter or set_table_close_kitchen_counter	close	kitchen_counter	5
set_table / settable	all six subtasks	mixed	all

Dataset-name inference maps clean pick/place datasets to pick,place and full set-table datasets to set_table.

Observation Format

Policy-facing keys emitted by the MS-HAB remote wrapper:

Key	Shape / dtype	Notes
task	str	Policy task string such as set table: pick object.
observation.state	(42,), float32	State vector.
observation.images.fetch_head	(3, 128, 128), float32	RGB image normalized to [0, 1].
observation.images.fetch_hand	(3, 128, 128), float32	RGB image normalized to [0, 1].

Depth images may exist in runtime observations, but the current policy-facing RGB contract filters out depth keys.

Action Format

Field	Value
Shape	(13,)
Dtype	float32
Range	[-1.0, 1.0]
Convention	mshab_normalized_controller_action

The action is validated by ActionSpec before being sent through the remote policy path.

Runtime Config

Pass the dedicated runtime interpreter through env_kwargs:

from praxis_eval import EvalConfig

config = EvalConfig(
    task="set_table",
    num_eval_per_task=3,
    num_parallel_env=1,
    output_dir="eval/mshab_set_table",
    env_kwargs={
        "python_bin": "/path/to/mshab-praxis/bin/python",
        "ms_asset_dir": "/data/mshab/maniskill_assets",
        "split": "val",
        "obs_mode": "rgb",
    },
)

Caveats

• Evaluation launches external subprocesses and communicates with the policy through praxis-remote.
• env.python_bin is required for normal evaluation unless the caller supplies an equivalent runtime path.
• Recording videos uses a dedicated single-env pass when enabled.
• The verifier defaults to a smaller random-action path; full evaluation uses mshab.praxis_eval from the runtime package.