How to randomize the initial state of the robot in simulation?

[stephane-caron]

The training script of the PPO balancer randomizes initial states in parallel environments, but the whole script altogether is hard to parse. What is a simple way to randomize the initial state of my Upkie in simulation?

[stephane-caron]

Here is a small balancer where we randomize the initial vertical position $z_0 \in [0.4, 1.2] \textrm{ m}$ and velocity $\dot{z}_0 \in [-0.1, 0.1] \textrm{ m/s}$:

import gymnasium as gym
import numpy as np
from upkie.utils.robot_state import RobotState, RobotStateRandomization

with gym.make(
    "UpkieGroundVelocity-v3",
    init_state=RobotState(
        position_base_in_world=np.array([0.0, 0.0, 0.8]),
        randomization=RobotStateRandomization(
            z=0.4,
            v_z=0.1,
        ),
    ),
) as env:
    env.reset()  # connects to the spine at first call
    action = np.zeros(env.action_space.shape)
    for step in range(1_000_000):
        observation, _, _, _, _ = env.step(action)
        pitch = observation[0]
        action[0] = 10.0 * pitch  # 1D action: [ground_velocity]