Environment Families

In CARL we provide several contextually extended environments where we can set and augment their contexts. With this we can analyze how the context and its augmentation influences the agent’s generalization capabilities, learning and behavior.

Although each environment has different tasks, goals and mechanics, the behavior of the dynamics and the rewards is influenced by physical properties.

CARL currently contains the following benchmarks which are contextually extended:

Classic Control

OpenAI gym classic control suite extended with several physics context features like gravity or friction

• CARL Classic Control Environments
  • CARL Pendulum Environment
  • CARL CartPole Environment
  • CARL Acrobot Environment
  • CARL MountainCar Environment

Box2D

OpenAI gym Box2D BipedalWalker, LunarLander and CarRacing, each with their own modification possibilities like new vehicles to race

• CARL Box 2D Environments
  • CARL BipedalWalker Environment
  • CARL LunarLander Environment
  • CARL VehicleRacing Environment

Brax

All Brax locomotion environments with exposed internal features like joint strength or torso mass

• CARL Brax Environments
  • CARL Ant Environment
  • CARL Fetch Environment
  • CARL Grasp Environment
  • CARL Halfcheetah Environment
  • CARL Humanoid Environment
  • CARL UR5e Environment

DeepMind Control Suite

Selected environment from the DeepMind Control Suite with controllable physics.

• CARL DMC Environments
  • CARL DMC Finger Environment
  • CARL DMC Fish Environment
  • CARL DMC Quadruped Environment
  • CARL DMC Walker Environment

Mario

Super Mario (TOAD-GAN), a procedurally generated jump’n’run game with control over level similarity

• CARL Mario Environment

RNA Design

RNADesign, an environment for RNA design given structure constraints with structures from different datasets to choose from

• CARL RNA Design Environment

Overview Table

Below every environment is listed with its number of context features and the type of action and observation space.

Env. Family	Name	# Context Features	Action Space	Obs. Space
classic_control	CARLMountainCarEnv	11	Discrete(3)	Box(-inf, inf, (13,), float32)
classic_control	CARLPendulumEnv	5	Box(-2.0, 2.0, (1,), float32)	Box(-inf, inf, (8,), float32)
classic_control	CARLAcrobotEnv	10	Discrete(3)	Box(-28.274333953857422, 282.74334716796875, (16,), float32)
classic_control	CARLCartPoleEnv	6	Discrete(2)	Box(-3.4028234663852886e+38, inf, (10,), float32)
classic_control	CARLMountainCarContinuousEnv	10	Box(-1.0, 1.0, (1,), float32)	Box(-inf, inf, (12,), float32)
box2d	CARLLunarLanderEnv	16	Discrete(4)	Box(-inf, inf, (24,), float32)
box2d	CARLVehicleRacingEnv	1	Box(-1.0, 1.0, (3,), float32)	Box(0, 255, (96, 96, 3), uint8)
box2d	CARLBipedalWalkerEnv	20	Box(-1.0, 1.0, (4,), float32)	Box(-inf, inf, (44,), float32)
brax	CARLAnt	7	Box(-1.0, 1.0, (8,), float32)	Box(-inf, inf, (94,), float32)
brax	CARLHalfcheetah	6	Box(-1.0, 1.0, (6,), float32)	Box(-inf, inf, (29,), float32)
brax	CARLHumanoid	5	Box(-1.0, 1.0, (17,), float32)	Box(-inf, inf, (304,), float32)
brax	CARLFetch	9	Box(-1.0, 1.0, (10,), float32)	Box(-inf, inf, (110,), float32)
brax	CARLGrasp	9	Box(-1.0, 1.0, (19,), float32)	Box(-inf, inf, (141,), float32)
brax	CARLUr5e	9	Box(-1.0, 1.0, (6,), float32)	Box(-inf, inf, (75,), float32)
dmc	CARLDmcFishEnv	14	Box(-1.0, 1.0, (5,), float64)	Box(-inf, inf, (24,), float32)
dmc	CARLDmcFingerEnv	18	Box(-1.0, 1.0, (2,), float64)	Box(-inf, inf, (9,), float32)
dmc	CARLDmcQuadrupedEnv	14	Box(-1.0, 1.1, (12,), float64)	Box(-inf, inf, (78,), float32)
dmc	CARLDmcWalkerEnv	14	Box(-1.0, 1.0, (6,), float64)	Box(-inf, inf, (24,), float32)
RNA	CARLRnaDesignEnv	5	Discrete(4)	Box(-inf, inf, (11,), float32)
Mario	CARLMarioEnv	4	Discrete(10)	Box(0, 255, (4, 64, 64), uint8)

previous

The CARL Environment

next

CARL Classic Control Environments