alive_bonus = 1.0
assert obs.ndim == 2 and action.ndim == 2
assert obs.shape == obs_next.shape and action.shape[0] == obs.shape[0]
vel = obs_next[:, 5]
ctrl_cost = 1e-3 * np.sum(np.square(action), axis=1)
reward = vel + alive_bonus - ctrl_cost
return np.minimum(np.maximum(-1000.0, reward), 1000.0)
if __name__ == "__main__":
env = HopperWrapper()
After Change
// obs = [cos(theta), sin(theta), dtheta/dt]
// To get the angle back from obs: atan2(sin(theta), cos(theta)).
theta = np.arctan2(
np.clip(obs[:, 1], -1.0, 1.0), np.clip(obs[:, 0], -1.0, 1.0))
// Do everything in (B,) space (single theta-, action- and
// reward values).
a = np.clip(action, -self.max_torque, self.max_torque)[0]
