gym-compete/gym_compete/new

from .multi_agent_env import MultiAgentEnv import numpy as np from gym import spaces import six class SumoEnv(MultiAgentEnv): ''' ''' def __init__(self, max_episode_steps=500, min_radius=1, max_radius=3.5, **kwargs): super(SumoEnv, self).__init__(**kwargs) self._max_episode_steps = max_episode_steps self._elapsed_steps = 0 self.GOAL_REWARD = 1000 self.RADIUS = self.MAX_RADIUS = self.current_max_radius = max_radius self.MIN_RADIUS = min_radius self.LIM_X = [(-2, 0), (0, 2)] self.LIM_Y = [(-2, 2), (-2, 2)] self.RANGE_X = self.LIM_X.copy() self.RANGE_Y = self.LIM_Y.copy() self.arena_id = self.env_scene.model.geom_names.index(six.b('arena')) self.arena_height = self.env_scene.model.geom_size[self.arena_id][1] * 2 self._set_geom_radius() self.agent_contacts = False def _past_limit(self): if self._max_episode_steps <= self._elapsed_steps: return True return False def _past_arena(self, agent_id): xy = self.agents[agent_id].get_qpos()[:2] r = np.sum(xy ** 2) ** 0.5 # print("Agent", agent_id, "at", r) if r > self.RADIUS: return True return False def _is_fallen(self, agent_id, limit=0.5): if self.agents[agent_id].team == 'ant': limit = 0.3 limit = limit + self.arena_height return bool(self.agents[agent_id].get_qpos()[2] <= limit) def _is_standing(self, agent_id, limit=0.9): limit = limit + self.arena_height return bool(self.agents[agent_id].get_qpos()[2] > limit) def get_agent_contacts(self): mjcontacts = self.env_scene.data._wrapped.contents.contact ncon = self.env_scene.model.data.ncon contacts = [] for i in range(ncon): ct = mjcontacts[i] g1 , g2 = ct.geom1, ct.geom2 g1 = self.env_scene.model.geom_names[g1] g2 = self.env_scene.model.geom_names[g2] if g1.find(six.b('agent')) >= 0 and g2.find(six.b('agent')) >= 0: if g1.find(six.b('agent0')) >= 0: if g2.find(six.b('agent1')) >= 0 and ct.dist < 0: contacts.append((g1, g2, ct.dist)) elif g1.find(six.b('agent1')) >= 0: if g2.find(six.b('agent0')) >= 0 and ct.dist < 0: contacts.append((g1, g2, ct.dist)) return contacts def goal_rewards(self, infos=None, agent_dones=None): self._elapsed_steps += 1 goal_rews = [0. for _ in range(self.n_agents)] fallen = [self._is_fallen(i) for i in range(self.n_agents)] timeup = self._past_limit() past_arena = [self._past_arena(i) for i in range(self.n_agents)] done = False agent_contacts = self.get_agent_contacts() if len(agent_contacts) > 0: # print('Detected contacts:', agent_contacts) self.agent_contacts = True if any(fallen): done = True for j in range(self.n_agents): if fallen[j]: print('Agent', j, 'fallen') goal_rews[j] -= self.GOAL_REWARD elif self.agent_contacts: goal_rews[j] += self.GOAL_REWARD infos[j]['winner'] = True # import ipdb; ipdb.set_trace() elif any(past_arena): done = True for j in range(self.n_agents): if past_arena[j]: print('Agent', j, 'past arena') goal_rews[j] -= self.GOAL_REWARD elif self.agent_contacts: goal_rews[j] += self.GOAL_REWARD infos[j]['winner'] = True elif timeup: for j in range(self.n_agents): goal_rews[j] -= self.GOAL_REWARD done = timeup or done return goal_rews, done def _set_observation_space(self): ob_spaces_limits = [] # nextra = 3 + self.n_agents - 1 nextra = 4 for i in range(self.n_agents): s = self.agents[i].observation_space.shape[0] h = np.ones(s+nextra) * np.inf l = -h ob_spaces_limits.append((l, h)) self.observation_space = spaces.Tuple( [spaces.Box(l, h) for l,h in ob_spaces_limits] ) def _get_obs(self): obs = [] dists = [] for i in range(self.n_agents): xy = self.agents[i].get_qpos()[:2] r = np.sqrt(np.sum(xy**2)) d = self.RADIUS - r # print(r, d) dists.append(d) for i in range(self.n_agents): ob = self.agents[i]._get_obs() mydist = np.asarray(dists[i]) if self.n_agents == 1: other_dist = np.asarray(self.RADIUS) elif self.n_agents == 2: other_dist = np.asarray(dists[1-i]) else: other_dist = np.asarray([dists[j] for j in range(self.n_agents) if j != i]) ob = np.concatenate( [ob.flat, np.asarray(self.RADIUS).flat, mydist.flat, other_dist.flat, np.asarray(self._max_episode_steps - self._elapsed_steps).flat ] ) obs.append(ob) return tuple(obs) def _reset_max_radius(self, version): decay_func_r = lambda x: 0.1 * np.exp(0.001 * x) vr = decay_func_r(version) self.current_max_radius = min(self.MAX_RADIUS, self.MIN_RADIUS + vr) # print(self.current_max_radius) def _reset_radius(self): self.RADIUS = np.random.uniform(self.MIN_RADIUS, self.current_max_radius) # print('setting Radus to', self.RADIUS) def _set_geom_radius(self): gs = self.env_scene.model.geom_size.copy() gs[self.arena_id][0] = self.RADIUS self.env_scene.model.__setattr__('geom_size', gs) self.env_scene.model.forward() def _reset_agents(self): # set agent 0 min_gap = 0.3 + self.MIN_RADIUS / 2 for i in range(self.n_agents): if i % 2 == 0: x = np.random.uniform(-self.RADIUS + min_gap, -0.3) y_lim = np.sqrt(self.RADIUS**2 - x**2) y = np.random.uniform(-y_lim + min_gap, y_lim - min_gap) else: x = np.random.uniform(0.3, self.RADIUS - min_gap) y_lim = np.sqrt(self.RADIUS**2 - x**2) y = np.random.uniform(-y_lim + min_gap, y_lim - min_gap) self.agents[i].set_xyz((x,y,None)) # print('setting agent', i, 'at', (x,y)) def _reset(self, version=None): self._elapsed_steps = 0 self.agent_contacts = False # self.RADIUS = self.START_RADIUS if version is not None: self._reset_max_radius(version) self._reset_radius() self._set_geom_radius() # print("here") _ = self.env_scene.reset() self._reset_agents() ob = self._get_obs() return ob def reset(self, margins=None, version=None): ob = self._reset(version=version) if margins: for i in range(self.n_agents): self.agents[i].set_margin(margins[i]) return ob

gym-compete/gym_compete/new_envs/sumo.py (167 lines of code) (raw):