Gymnasium Robotics

3D contact-rich manipulation suites from Gymnasium Robotics.

Description

A collection of 3D contact-rich manipulation environments built using the Gymnasium Robotics API. The environments use the MuJoCo physics engine to simulate diverse robotic kinematics, explicitly tracking continuous-control primitives like reaching, pushing, sliding, and pick-and-place dynamics.

import stable_worldmodel as swm

# Reach environment (pure gripper kinematics)
world = swm.World('swm/FetchReach-v3', num_envs=4, image_shape=(224, 224))

# Contact-rich block manipulation environments
world = swm.World('swm/FetchPush-v3', num_envs=4, image_shape=(224, 224))
world = swm.World('swm/FetchSlide-v3', num_envs=4, image_shape=(224, 224))
world = swm.World('swm/FetchPickAndPlace-v3', num_envs=4, image_shape=(224, 224))

Available Environments

Environment	Environment ID	Task Objective
Fetch Reach	`swm/FetchReach-v3`	Move the gripper to a coordinate in mid-air
Fetch Push	`swm/FetchPush-v3`	Push a block to a stationary coordinate
Fetch Slide	`swm/FetchSlide-v3`	Strike a block so it slides across a slippery table
Fetch Pick And Place	`swm/FetchPickAndPlace-v3`	Grasp a block and lift it to a mid-air coordinate

Fetch Manipulation Suite

An agent controls a 7-DoF Fetch robotic arm. The agent manipulates explicit Cartesian coordinates to move the gripper and actuate the fingers to interact with the environment.

Success criteria: The gripper (in Reach) or the physical block (in Push, Slide, Pick And Place) must be moved within a short euclidean distance threshold of the goal state coordinates.

Environment Specs

Property	Value
Action Space	`Box(-1, 1, shape=(4,))` — 3D Cartesian velocity + Gripper control
Observation Space (Reach)	`Box(-inf, inf, shape=(13,))` — Flattened `observation` + `desired_goal` arrays
Observation Space (Push/Slide/Pick)	`Box(-inf, inf, shape=(28,))` — Flattened `observation` + `desired_goal` arrays
Reward	Standard sparse/dense task rewards depending on upstream configurations
Episode Length	50 steps (default)
Render Size	Configurable via `resolution=224` on init
Physics	MuJoCo

Task Types

`env_name`	Manipulated Objects	Task Dynamics
`FetchReach-v3`	None	Pure kinematic arm movement
`FetchPush-v3`	1 Block	Surface contact physics
`FetchSlide-v3`	1 Block	Momentum and low-friction sliding
`FetchPickAndPlace-v3`	1 Block	Grasping and mid-air suspension

Info Dictionary

The info dict returned by step() and reset() is strictly formatted to standard SWM API integrations:

Key	Shape (Push/Pick/Slide)	Shape (Reach)	Description
`env_name`	`str`	`str`	The ID of the environment (e.g. `FetchPush-v4`)
`state`	`(28,)`	`(13,)`	Full flattened state array natively concatenating `observation` + `desired_goal`
`proprio`	`(25,)`	`(10,)`	Isolated agent internal state (gripper poses, joint velocities, object physics)
`goal_state`	`(3,)`	`(3,)`	The 3D Cartesian XYZ target coordinates for the objective

Variation Space

The environment supports extensive domain randomization across both visual textures and explicitly intercepting internal MuJoCo physics states.

Factor	Type	Description
`table.color`	RGBBox	Table surface color (default: `[0.3, 0.3, 0.3]`)
`object.color`	RGBBox	Manipulated block color (default: `[0.8, 0.1, 0.1]`)
`background.color`	RGBBox	Studio/Skybox background color (default: `[0.1, 0.1, 0.1]`)
`light.intensity`	Box(0.0, 1.0)	Diffuse scene lighting intensity (default: `0.7`)
`camera.angle_delta`	Box(-10.0, 10.0, shape=(1, 2))	Azimuth/elevation camera perturbations (degrees)
`agent.start_position`	Box([1.25, 0.6], [1.45, 0.9])	Starting 2D (x,y) coordinates for the initial gripper spawn targeting
`block.start_position`	Box([1.15, 0.6], [1.45, 0.9])	Explicit 2D (x,y) override intercepting initial qpos spawning
`block.angle`	Box(-π, π)	Explicit Z-rotation override intercepting initial qpos quaternions
`block.mass`	Box(0.01, 50.0)	Mass of the manipulated block (kg); default inherits the MuJoCo model's `body_mass`
`goal.start_position`	Box([1.15, 0.6, 0.424], [1.45, 0.9, 0.424])	Explicit XYZ override redefining visual and reward goal markers
`rendering.transparent_arm`	Discrete(2)	When set to `1`, natively lowers the PyMuJoCo alpha channels mapping to the robot body, making the robot translucent (see-through)

block.start_position, block.angle, and block.mass are only present in environments with a manipulated object (FetchPush-v3, FetchSlide-v3, FetchPickAndPlace-v3). FetchReach-v3 exposes the remaining 8 factors.

Default Configuration

By default, the following visual factors are natively tracked and continuously randomized upon resetting if no variations are specified:

table.color
object.color
light.intensity
background.color
camera.angle_delta

Overriding Physics and Randomization

To randomize physical spawn properties or enable fully deterministic data generation workflows, simply pass the bounding properties via the options array natively supported by the Gymnasium wrapper during environment reset:

# Randomize environment visuals AND strictly dictate starting block positions
obs, info = world.reset(options={
    'variation': [
        'table.color', 'object.color', 'background.color', 
        'agent.start_position', 'block.start_position'
    ]
})

# Or bypass randomization altogether and strictly force perfect coordinate reproducibility:
obs, info = world.reset(options={
    'variation_values': {
        'block.start_position': [1.3, 0.7],
        'goal.start_position': [1.3, 0.7, 0.4247]
    }
})