CARL Brax Environments

In CARL all Brax locomotion environments are included. As context features there are external features like gravity or friction or internal features like joint strength or torso mass.

CARL Ant Environment

Here the agent needs to learn how to control a four-legged ant in order to run (quickly) into a given direction.

Defaults and Bounds

Context Feature	Default	Bounds
joint_stiffness	5000.0	(1, inf, <class ‘float’>)
gravity	-9.8	(-inf, -0.1, <class ‘float’>)
friction	0.6	(-inf, inf, <class ‘float’>)
angular_damping	-0.05	(-inf, inf, <class ‘float’>)
actuator_strength	300.0	(1, inf, <class ‘float’>)
joint_angular_damping	35.0	(0, inf, <class ‘float’>)
torso_mass	10.0	(0.1, inf, <class ‘float’>)

CARL Fetch Environment

Fetch trains a robotic dog to run to a target location. The target radius and distance as well as physical properties can be varied via the context features.

Defaults and Bounds

Context Feature	Default	Bounds
joint_stiffness	5000.0	(1, inf, <class ‘float’>)
gravity	-9.8	(-inf, -0.1, <class ‘float’>)
friction	0.6	(-inf, inf, <class ‘float’>)
angular_damping	-0.05	(-inf, inf, <class ‘float’>)
actuator_strength	300.0	(1, inf, <class ‘float’>)
joint_angular_damping	35.0	(0, inf, <class ‘float’>)
torso_mass	1.0	(0.1, inf, <class ‘float’>)
target_radius	2.0	(0.1, inf, <class ‘float’>)
target_distance	15.0	(0.1, inf, <class ‘float’>)

CARL Grasp Environment

In CARL Grasp the agent is trained to pick up an object with a robot hand. Three bodies are observed by Grasp: ‘Hand’, ‘Object’, and ‘Target’. When Object reaches Target, the agent is rewarded. Apart from Grasp’s pyhiscal properties the target radius, height and distance are also varied.

Defaults and Bounds

Context Feature	Default	Bounds
joint_stiffness	5000.0	(1, inf, <class ‘float’>)
gravity	-9.8	(-inf, -0.1, <class ‘float’>)
friction	0.6	(-inf, inf, <class ‘float’>)
angular_damping	-0.05	(-inf, inf, <class ‘float’>)
actuator_strength	300.0	(1, inf, <class ‘float’>)
joint_angular_damping	50.0	(0, inf, <class ‘float’>)
target_radius	1.1	(0.1, inf, <class ‘float’>)
target_distance	10.0	(0.1, inf, <class ‘float’>)
target_height	8.0	(0.1, inf, <class ‘float’>)

CARL Halfcheetah Environment

A Halfcheetah is trained to run in a given direction. The context features can vary physical properties.

Defaults and Bounds

Context Feature	Default	Bounds
joint_stiffness	15000.0	(1, inf, <class ‘float’>)
gravity	-9.8	(-inf, -0.1, <class ‘float’>)
friction	0.6	(-inf, inf, <class ‘float’>)
angular_damping	-0.05	(-inf, inf, <class ‘float’>)
joint_angular_damping	20.0	(0, inf, <class ‘float’>)
torso_mass	9.457333	(0.1, inf, <class ‘float’>)

CARL Humanoid Environment

Here, a Humanoid needs to learn how to walk forward.

Defaults and Bounds

Context Feature	Default	Bounds
gravity	-9.8	(-inf, -0.1, <class ‘float’>)
friction	0.6	(-inf, inf, <class ‘float’>)
angular_damping	-0.05	(-inf, inf, <class ‘float’>)
joint_angular_damping	20.0	(0, inf, <class ‘float’>)
torso_mass	8.907463	(0.1, inf, <class ‘float’>)

CARL UR5e Environment

The agent needs to learn how to move a ur5e robot arm and its end effector to a sequence of targets. The robot arm has 6 joints.

Defaults and Bounds

Context Feature	Default	Bounds
joint_stiffness	40000.0	(1, inf, <class ‘float’>)
gravity	-9.81	(-inf, -0.1, <class ‘float’>)
friction	0.6	(-inf, inf, <class ‘float’>)
angular_damping	-0.05	(-inf, inf, <class ‘float’>)
actuator_strength	100.0	(1, inf, <class ‘float’>)
joint_angular_damping	50.0	(0, 360, <class ‘float’>)
target_radius	0.02	(0.01, inf, <class ‘float’>)
target_distance	0.5	(0.01, inf, <class ‘float’>)
torso_mass	1.0	(0, inf, <class ‘float’>)

previous

CARL Box 2D Environments

next

CARL DMC Environments