Basic Usage Example

The environment can work out of the box without too much customization.

A gym env could be instantiated without any extra arguments. By default, it spawns two agents in the Vegas (IROS 2020) map. You can find the image of the map at gym/f110_gym/envs/maps/vegas.png. At instantiation, the index of the ego agent in the list of agents is 0.

The agents can be reset by calling the reset() method using a numpy ndarray of size (num_agents, 2), where each row represents an agent, and the columns are the (x, y) coordinate of each agent.

The reset() and step() method returns:

An observation dictionary
A step reward, which in the current release is the physics timestep used.
A done boolean indicator, flips to true when either a collision happens or the ego agent finishes 2 laps.
An info dictionary. Empty in the current release.

The action taken by the step() function is a numpy ndarray of size (num_agents, 2), where each row represents an agent’s action (indices corresponds to the list of agents), and the columns are control inputs (steering angle, velocity).

A working example can be found in examples/waypoint_follow.py.

The following pseudo code provides a skeleton for creating a simulation loop.

import gym
import numpy as np
from your_custom_policy import planner # the policy/motion planner that you create

# instantiating the environment
racecar_env = gym.make('f110_gym:f110-v0')
obs, step_reward, done, info = racecar_env.reset(np.array([[0., 0., 0.], # pose of ego
                                                           [2., 0., 0.]])) # pose of 2nd agent
# instantiating your policy
planner = planner()

# simulation loop
lap_time = 0.

# loops when env not done
while not done:
    # get action based on the observation
    actions = planner.plan(obs)

    # stepping through the environment
    obs, step_reward, done, info = racecar_env.step(actions)

    lap_time += step_reward

For a more in-depth example that provides more customization to the environment, see Customized Usage Example.