Gymnasium/gym/envs/mujoco/walker2d.py

import numpy as np

from gym import utils
from gym.envs.mujoco import MuJocoPyEnv
from gym.spaces import Box


class Walker2dEnv(MuJocoPyEnv, utils.EzPickle):
    metadata = {
        "render_modes": [
            "human",
            "rgb_array",
            "depth_array",
            "single_rgb_array",
            "single_depth_array",
        ],
        "render_fps": 125,
    }

    def __init__(self, **kwargs):
        observation_space = Box(low=-np.inf, high=np.inf, shape=(17,), dtype=np.float64)
        MuJocoPyEnv.__init__(
            self, "walker2d.xml", 4, observation_space=observation_space, **kwargs
        )
        utils.EzPickle.__init__(self, **kwargs)

    def step(self, a):
        posbefore = self.sim.data.qpos[0]
        self.do_simulation(a, self.frame_skip)
        posafter, height, ang = self.sim.data.qpos[0:3]

        self.renderer.render_step()

        alive_bonus = 1.0
        reward = (posafter - posbefore) / self.dt
        reward += alive_bonus
        reward -= 1e-3 * np.square(a).sum()
        terminated = not (height > 0.8 and height < 2.0 and ang > -1.0 and ang < 1.0)
        ob = self._get_obs()
        return ob, reward, terminated, False, {}

    def _get_obs(self):
        qpos = self.sim.data.qpos
        qvel = self.sim.data.qvel
        return np.concatenate([qpos[1:], np.clip(qvel, -10, 10)]).ravel()

    def reset_model(self):
        self.set_state(
            self.init_qpos
            + self.np_random.uniform(low=-0.005, high=0.005, size=self.model.nq),
            self.init_qvel
            + self.np_random.uniform(low=-0.005, high=0.005, size=self.model.nv),
        )
        return self._get_obs()

    def viewer_setup(self):
        assert self.viewer is not None
        self.viewer.cam.trackbodyid = 2
        self.viewer.cam.distance = self.model.stat.extent * 0.5
        self.viewer.cam.lookat[2] = 1.15
        self.viewer.cam.elevation = -20