Pendulum

A simple inverted pendulum environment.

class jax_agents.environments.pendulum.PendulumEnv(seed=1996)

Bases: object

The pendulum starts almost upright, the goal is to prevent falling down.

States:

Type: Box(4) Num State Min Max 0 Pendulum Angle 0(downward) 2 pi 2 Pendulum Angular Velocity -Inf Inf

Actions:

Type: Box(1) Num Action Min Max 0 Torque -4.0 4.0

Orientation:

angle: pi/2 -> to the right angular velocity: positive -> counterclockwise

Reward:

Reward is 1 for every step in the upright position, -0.1 otherwise

Starting State:

Angle = pi + (-10, +10) (i.e. pointing upwards) Velocity = 0.0

Episode Length:

10 seconds, i.e. 200 steps at 20 Hz

check_if_done(state)

Check if episode needs to restart.

initialize_rendering(log, fig)

Initialize Matplotlib axes for rendering.

norm_state(state)

Norm the state to have zero mean and unit variance.

pendulum_dynamics(state, action)

Environment’s dynamics. Return the derivative of the state.

render(log)

Render the given timestep.

rescale_action(scaled_action)

Rescale the action from the policy (which are between -1 and 1).

reset()

Reset the state to start a new episode.

reward_func(state, action, next_state)

Return the reward for the given transition.

step(state, action)

Integrate the dynamics on step forward.