CARL DMC Environments

CARL includes the Finger, Fish, Quadruped and Walker environments from the DeepMind Control Suite. The context features control the MuJoCo physics engine, e.g. the floor friction.

CARL DMC Finger Environment

Screenshot of CARLDmcFinger

The agent needs to learn to spin an object using the finger.

Defaults and Bounds

Context Feature

Default

Bounds

gravity

-9.81

(-inf, -0.1, <class ‘float’>)

friction_tangential

1

(0, inf, <class ‘float’>)

friction_torsional

1

(0, inf, <class ‘float’>)

friction_rolling

1

(0, inf, <class ‘float’>)

timestep

0.004

(0.001, 0.1, <class ‘float’>)

joint_damping

1.0

(0, inf, <class ‘float’>)

joint_stiffness

0.0

(0, inf, <class ‘float’>)

actuator_strength

1

(0, inf, <class ‘float’>)

density

5000.0

(0, inf, <class ‘float’>)

viscosity

0.0

(0, inf, <class ‘float’>)

geom_density

1.0

(0, inf, <class ‘float’>)

wind_x

0.0

(-inf, inf, <class ‘float’>)

wind_y

0.0

(-inf, inf, <class ‘float’>)

wind_z

0.0

(-inf, inf, <class ‘float’>)

limb_length_0

0.17

(0.01, 0.2, <class ‘float’>)

limb_length_1

0.16

(0.01, 0.2, <class ‘float’>)

spinner_radius

0.04

(0.01, 0.05, <class ‘float’>)

spinner_length

0.18

(0.01, 0.4, <class ‘float’>)

CARL DMC Fish Environment

Screenshot of CARLDmcFish

In Fish, the agent needs to swim as a simulated fish.

Defaults and Bounds

Context Feature

Default

Bounds

gravity

-9.81

(-inf, -0.1, <class ‘float’>)

friction_tangential

1

(0, inf, <class ‘float’>)

friction_torsional

1

(0, inf, <class ‘float’>)

friction_rolling

1

(0, inf, <class ‘float’>)

timestep

0.004

(0.001, 0.1, <class ‘float’>)

joint_damping

1.0

(0, inf, <class ‘float’>)

joint_stiffness

0.0

(0, inf, <class ‘float’>)

actuator_strength

1

(0, inf, <class ‘float’>)

density

5000.0

(0, inf, <class ‘float’>)

viscosity

0.0

(0, inf, <class ‘float’>)

geom_density

1.0

(0, inf, <class ‘float’>)

wind_x

0.0

(-inf, inf, <class ‘float’>)

wind_y

0.0

(-inf, inf, <class ‘float’>)

wind_z

0.0

(-inf, inf, <class ‘float’>)

CARL DMC Quadruped Environment

Screenshot of CARLDmcQuadruped

The agent’s goal is to walk efficiently with the quadruped robot. Influence of context settings on an agent trained on the default environment:

Influence of context settings on an agent trained on the default environment.
Defaults and Bounds

Context Feature

Default

Bounds

gravity

-9.81

(-inf, -0.1, <class ‘float’>)

friction_tangential

1

(0, inf, <class ‘float’>)

friction_torsional

1

(0, inf, <class ‘float’>)

friction_rolling

1

(0, inf, <class ‘float’>)

timestep

0.005

(0.001, 0.1, <class ‘float’>)

joint_damping

1.0

(0, inf, <class ‘float’>)

joint_stiffness

0.0

(0, inf, <class ‘float’>)

actuator_strength

1

(0, inf, <class ‘float’>)

density

0.0

(0, inf, <class ‘float’>)

viscosity

0.0

(0, inf, <class ‘float’>)

geom_density

1.0

(0, inf, <class ‘float’>)

wind_x

0.0

(-inf, inf, <class ‘float’>)

wind_y

0.0

(-inf, inf, <class ‘float’>)

wind_z

0.0

(-inf, inf, <class ‘float’>)

CARL DMC Walker Environment

Screenshot of CARLDmcWalker

The walker robot is supposed to move forward as fast as possible. Influence of context settings on an agent trained on the default environment:

Influence of context settings on an agent trained on the default environment.
Defaults and Bounds

Context Feature

Default

Bounds

gravity

-9.81

(-inf, -0.1, <class ‘float’>)

friction_tangential

1

(0, inf, <class ‘float’>)

friction_torsional

1

(0, inf, <class ‘float’>)

friction_rolling

1

(0, inf, <class ‘float’>)

timestep

0.0025

(0.001, 0.1, <class ‘float’>)

joint_damping

1.0

(0, inf, <class ‘float’>)

joint_stiffness

0.0

(0, inf, <class ‘float’>)

actuator_strength

1

(0, inf, <class ‘float’>)

density

5000.0

(0, inf, <class ‘float’>)

viscosity

0.0

(0, inf, <class ‘float’>)

geom_density

1.0

(0, inf, <class ‘float’>)

wind_x

0.0

(-inf, inf, <class ‘float’>)

wind_y

0.0

(-inf, inf, <class ‘float’>)

wind_z

0.0

(-inf, inf, <class ‘float’>)