Implementation:Haosulab ManiSkill GymUtils

Knowledge Sources	Haosulab_ManiSkill
Domains	Robotics, Simulation, Gymnasium Integration
Last Updated	2026-02-15 08:00 GMT

Overview

Concrete tool for Gymnasium/OpenAI Gym utilities including observation/action space manipulation, scalar extraction, and action normalization.

Description

The gym_utils.py module provides various utility functions for working with Gymnasium observation and action spaces, commonly needed when integrating ManiSkill environments with reinforcement learning frameworks.

Environment utilities:

• find_max_episode_steps_value(env) -- Traverses the wrapper chain to find the max_episode_steps parameter, handling SyncVectorEnv, ManiSkillVectorEnv, and standard wrappers.

Scalar extraction:

• extract_scalars_from_info(info, blacklist, batch_size) -- Recursively extracts scalar metrics from info dictionaries returned by env.step(). Supports both single and batched environments.

Action normalization:

• clip_and_scale_action(action, low, high) -- Clips actions to [-1, 1] and scales to [low, high] range.
• inv_scale_action(action, low, high) -- Inverse of clip_and_scale_action (without clipping).
• normalize_action_space(action_space) -- Converts a Box action space to normalized [-1, 1] range.

Type utilities:

• get_dtype_bounds(dtype) -- Gets min/max values for numpy dtypes (floating, integer, bool).
• convert_observation_to_space(observation, prefix, unbatched) -- Recursively converts a sample observation (dict, array, scalar) into a corresponding Gymnasium observation space.

Version detection:

• IS_GYMNASIUM_1 -- Boolean flag detecting if gymnasium version > 1.0.0.

Usage

Import from mani_skill.utils.gym_utils when building training pipelines, converting observation spaces, or normalizing actions for RL algorithms.

Code Reference

Source Location

• Repository: Haosulab_ManiSkill
• File: mani_skill/utils/gym_utils.py

Signature

def find_max_episode_steps_value(env) -> Optional[int]: ...
def extract_scalars_from_info(info: dict, blacklist=(), batch_size=1) -> dict[str, float]: ...
def clip_and_scale_action(action, low, high) -> torch.Tensor: ...
def inv_scale_action(action, low, high) -> np.ndarray: ...
def normalize_action_space(action_space: spaces.Box) -> spaces.Box: ...
def get_dtype_bounds(dtype: np.dtype) -> tuple: ...
def convert_observation_to_space(observation, prefix="", unbatched=False) -> spaces.Space: ...

Import

from mani_skill.utils.gym_utils import (
    clip_and_scale_action,
    convert_observation_to_space,
    extract_scalars_from_info,
)

I/O Contract

Inputs

Name	Type	Required	Description
action	torch.Tensor	Yes	Action tensor in [-1, 1] range
low	float/np.ndarray	Yes	Lower bound of action range
high	float/np.ndarray	Yes	Upper bound of action range
observation	dict/np.ndarray	Yes	Sample observation for space inference

Outputs

Name	Type	Description
scaled_action	torch.Tensor	Action scaled to [low, high] range
space	spaces.Space	Inferred Gymnasium observation space
scalars	dict[str, float]	Extracted scalar metrics

Usage Examples

Basic Usage

import torch
from mani_skill.utils.gym_utils import clip_and_scale_action, extract_scalars_from_info

# Scale normalized actions to joint limits
action = torch.tensor([0.5, -0.3])
scaled = clip_and_scale_action(action, low=-3.14, high=3.14)

# Extract scalar metrics from info dict
info = {"success": True, "dist": 0.05, "nested": {"metric": 1.0}}
scalars = extract_scalars_from_info(info)
# {"success": 1.0, "dist": 0.05, "nested.metric": 1.0}

Related Pages

• Environment:Haosulab_ManiSkill_Python_SAPIEN_Core
• Haosulab_ManiSkill_CPUGymWrapper