def __init__(self, env, device='cpu', use_td3=True):
self.env = env
self.device = device
self.is_discrete_action = isinstance(env.action_space, gym.spaces.discrete.Discrete)
self.obs_dim = env.observation_space.shape[0]
gamma = 0.99
self.gamma = gamma
if self.is_discrete_action:
self.num_act = env.action_space.n
self.algo = DQNWithRewardAlgo(self.obs_dim, self.num_act, gamma, use_td3=use_td3, device=device)
else:
self.act_dim = env.action_space.shape[0]
self.algo = DDPGAlgo(self.obs_dim, self.act_dim, gamma, use_td3=use_td3, device=device)
self.replay_buffer = ReplayBuffer()
self.planner = Planner(trans_fn=self.planner_trans_fn, use_td3=use_td3, gamma=gamma, device=device)
self.estimate_std()
def __init__(self, config_module=None):
"""
param:config_module: a scenario/robot specific module to prepare setup,
that has the following members:
get_all_conditions() -> return a list of conditions
get_all_actions() -> return a list of actions
"""
self.memory = Memory()
self.worldstate = WorldState()
self.actions = set()
if config_module is not None:
for condition in config_module.get_all_conditions(self.memory):
Condition.add(condition)
for action in config_module.get_all_actions(self.memory):
self.actions.add(action)
self.planner = Planner(self.actions, self.worldstate, None)
self._last_goal = None
self._preempt_requested = False # preemption mechanism
def test_planner():
assert Planner()