Gymnasium/gym/envs/mujoco/pusher.py

import numpy as np

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


class PusherEnv(mujoco_env.MujocoEnv, utils.EzPickle):
    metadata = {
        "render_modes": [
            "human",
            "rgb_array",
            "depth_array",
            "single_rgb_array",
            "single_depth_array",
        ],
        "render_fps": 20,
    }

    def __init__(self, **kwargs):
        utils.EzPickle.__init__(self)
        observation_space = Box(low=-np.inf, high=np.inf, shape=(23,), dtype=np.float64)
        mujoco_env.MujocoEnv.__init__(
            self,
            "pusher.xml",
            5,
            mujoco_bindings="mujoco_py",
            observation_space=observation_space,
            **kwargs
        )

    def step(self, a):
        vec_1 = self.get_body_com("object") - self.get_body_com("tips_arm")
        vec_2 = self.get_body_com("object") - self.get_body_com("goal")

        reward_near = -np.linalg.norm(vec_1)
        reward_dist = -np.linalg.norm(vec_2)
        reward_ctrl = -np.square(a).sum()
        reward = reward_dist + 0.1 * reward_ctrl + 0.5 * reward_near

        self.do_simulation(a, self.frame_skip)

        self.renderer.render_step()

        ob = self._get_obs()
        done = False
        return ob, reward, done, dict(reward_dist=reward_dist, reward_ctrl=reward_ctrl)

    def viewer_setup(self):
        assert self.viewer is not None
        self.viewer.cam.trackbodyid = -1
        self.viewer.cam.distance = 4.0

    def reset_model(self):
        qpos = self.init_qpos

        self.goal_pos = np.asarray([0, 0])
        while True:
            self.cylinder_pos = np.concatenate(
                [
                    self.np_random.uniform(low=-0.3, high=0, size=1),
                    self.np_random.uniform(low=-0.2, high=0.2, size=1),
                ]
            )
            if np.linalg.norm(self.cylinder_pos - self.goal_pos) > 0.17:
                break

        qpos[-4:-2] = self.cylinder_pos
        qpos[-2:] = self.goal_pos
        qvel = self.init_qvel + self.np_random.uniform(
            low=-0.005, high=0.005, size=self.model.nv
        )
        qvel[-4:] = 0
        self.set_state(qpos, qvel)
        return self._get_obs()

    def _get_obs(self):
        return np.concatenate(
            [
                self.sim.data.qpos.flat[:7],
                self.sim.data.qvel.flat[:7],
                self.get_body_com("tips_arm"),
                self.get_body_com("object"),
                self.get_body_com("goal"),
            ]
        )