def _create_MPPI_controller(self):
ctrl = mppi.MPPI(
dynamics=self._predict,
running_cost=self.cost_fn,
nx=self.state_dim,
noise_sigma=self.noise_sigma,
num_samples=self.n_samples,
horizon=self.horizon,
lambda_=self.lambda_,
device=TORCH_DEVICE,
u_min=torch.tensor(self.action_lb,
dtype=torch.float,
device=TORCH_DEVICE),
u_max=torch.tensor(self.action_ub,
dtype=torch.float,
device=TORCH_DEVICE),
)
return ctrl
action = np.random.uniform(low=ACTION_LOW, high=ACTION_HIGH)
env.step([action])
# env.render()
new_data[i, :nx] = pre_action_state
new_data[i, nx:] = action
train(new_data)
logger.info("bootstrapping finished")
env = wrappers.Monitor(env, '/tmp/mppi/', force=True)
env.reset()
if downward_start:
env.env.state = [np.pi, 1]
mppi_gym = mppi.MPPI(dynamics,
running_cost,
nx,
noise_sigma,
num_samples=N_SAMPLES,
horizon=TIMESTEPS,
lambda_=lambda_,
device=d,
u_min=torch.tensor(ACTION_LOW,
dtype=torch.double,
device=d),
u_max=torch.tensor(ACTION_HIGH,
dtype=torch.double,
device=d))
total_reward, data = mppi.run_mppi(mppi_gym, env, train)
logger.info("Total reward %f", total_reward)
theta_dt = state[:, 1]
action = action[:, 0]
cost = angle_normalize(
theta)**2 + 0.1 * theta_dt**2 + 0.001 * action**2
return cost
def train(new_data):
pass
downward_start = True
env = gym.make(ENV_NAME).env # bypass the default TimeLimit wrapper
env.reset()
if downward_start:
env.state = [np.pi, 1]
env = wrappers.Monitor(env, '/tmp/mppi/', force=True)
env.reset()
if downward_start:
env.env.state = [np.pi, 1]
nx = 2
mppi_gym = mppi.MPPI(dynamics,
running_cost,
nx,
noise_sigma,
num_samples=N_SAMPLES,
horizon=TIMESTEPS,
lambda_=lambda_)
total_reward = mppi.run_mppi(mppi_gym, env, train)
logger.info("Total reward %f", total_reward)