def get_state_reward(self, fw): """Compute the state reward.""" state_r = feature.feature(lambda t, x, u: 0.0) for lane, w_lane in zip(self.world.lanes, self.w_lanes): if self.is_human: lane_gaussian_std = constants.LANE_REWARD_STDEV_h else: lane_gaussian_std = constants.LANE_REWARD_STDEV_r state_r += w_lane * lane.gaussian(fw=fw, stdev=lane_gaussian_std) for fence, w_fence in zip(self.world.fences, self.w_fences): if self.fence_sigmoid: # sigmoid fence reward state_r += w_fence * fence.sigmoid(fw=fw) else: # gaussian-shaped fence reward state_r += w_fence * fence.gaussian(fw=fw) if self.speed is not None: state_r += self.w_speed * feature.speed(self.speed) for other_traj, w_other_traj in zip(self.other_car_trajs, self.w_other_car_trajs): if self.fine_behind: state_r += (w_other_traj * other_traj.gaussian(fw, length=.14, width=.03)) else: state_r += (w_other_traj * other_traj.gaussian(fw, length=.14, width=.03) + other_traj.not_behind(fw, self.w_behind)) for other_truck_traj, w_other_truck_traj in zip( self.other_truck_trajs, self.w_other_truck_trajs): state_r += (w_other_truck_traj * other_truck_traj.sigmoid(fw)) return state_r
def simple_reward(self, trajs=None, lanes=None, roads=None, fences=None, speed=1., speed_import=1.): if lanes is None: lanes = self.lanes if roads is None: roads = self.roads if fences is None: fences = self.fences if trajs is None: trajs = [c.linear for c in self.cars] elif isinstance(trajs, car.Car): trajs = [c.linear for c in self.cars if c != trajs] r = 0.1 * feature.control() theta = [1., -50., 10., 10., -60.] # Simple model # theta = [.959, -46.271, 9.015, 8.531, -57.604] for lane in lanes: r = r + theta[0] * lane.gaussian() for fence in fences: r = r + theta[1] * fence.gaussian() for road in roads: r = r + theta[2] * road.gaussian(10.) if speed is not None: r = r + speed_import * theta[3] * feature.speed(speed) for traj in trajs: r = r + theta[4] * traj.gaussian() return r
def simple_reward(self, trajs=None, lanes=None, roads=None, fences=None, speed=1., speed_import=1.): if lanes is None: lanes = self.lanes if roads is None: roads = self.roads if fences is None: fences = self.fences if trajs is None: trajs = [c.linear for c in self.cars] elif isinstance(trajs, car.Car): trajs = [c.linear for c in self.cars if c!=trajs] r = 0.1*feature.control() theta = [1., -50., 10., 10., -60.] # Simple model # theta = [.959, -46.271, 9.015, 8.531, -57.604] for lane in lanes: r = r+theta[0]*lane.gaussian() for fence in fences: r = r+theta[1]*fence.gaussian() for road in roads: r = r+theta[2]*road.gaussian(10.) if speed is not None: r = r+speed_import*theta[3]*feature.speed(speed) for traj in trajs: r = r+theta[4]*traj.gaussian() return r
def feature_calc(t, x, u, target_car): res = [0, 0, 0, 0, 0] for lane in self.lanes: res[0] = res[0] + lane.gaussian()(t, x, u).eval() for fence in self.fences: res[1] = res[1] + fence.gaussian()(t, x, u).eval() for road in self.roads: res[2] = res[2] + road.gaussian(10.)(t, x, u).eval() res[3] = res[3] + feature.speed(1.)(t, x, u) for car in self.cars: if car != target_car: res[4] = res[4] + car.traj.gaussian()(t, x, u).eval() return res
def state_rewards(self, fw): """Compute the individual state rewards and return them as a dictionary with keys that describe the rewards.""" rewards = {} state_r = feature.feature(lambda t, x, u: 0.0) for i, (lane, w_lane) in enumerate(zip(self.world.lanes, self.w_lanes)): if self.is_human: lane_gaussian_std = constants.LANE_REWARD_STDEV_h else: lane_gaussian_std = constants.LANE_REWARD_STDEV_r rewards['lane gaussian ' + str(i)] = w_lane * lane.gaussian( fw=fw, stdev=lane_gaussian_std) for i, (fence, w_fence) in enumerate(zip(self.world.fences, self.w_fences)): if self.fence_sigmoid: # sigmoid fence reward rewards['fence sigmoid ' + str(i)] = w_fence * fence.sigmoid(fw=fw) else: # gaussian-shaped fence reward rewards['fence gaussian ' + str(i)] = w_fence * fence.gaussian(fw=fw) if self.speed is not None: rewards['speed'] = self.w_speed * feature.speed(self.speed) for i, (other_car_traj, w_other_car_traj) in enumerate( zip(self.other_car_trajs, self.w_other_car_trajs)): if self.is_human: w = w_other_car_traj else: w = w_other_car_traj if self.fine_behind: rewards['other traj gaussian ' + str(i)] = ( w * other_car_traj.gaussian(fw, length=.1, width=.03)) else: rewards['other traj gaussian ' + str(i)] = ( w * other_car_traj.gaussian(fw, length=.14, width=.03)) rewards['other traj not behind ' + str(i)] = other_car_traj.not_behind(fw, self.w_behind) for i, (other_truck_traj, w_other_truck_traj) in enumerate( zip(self.other_truck_trajs, self.w_other_truck_trajs)): rewards['other truck sigmoid ' + str(i)] = (w_other_truck_traj * other_truck_traj.sigmoid(fw)) return rewards
def simple_reward(self, trajs=None, lanes=None, roads=None, fences=None, speed=1., speed_import=1.): # skapar simple reward for en bil if lanes is None: lanes = self.lanes if roads is None: roads = self.roads if fences is None: fences = self.fences if trajs is None: trajs = [c.linear for c in self.cars] elif isinstance(trajs, car.Car): trajs = [c.linear for c in self.cars if c != trajs] elif isinstance(trajs, static_obj.Car): trajs = [c.linear for c in self.cars if c != trajs] r = 0.1 * feature.control() theta = [1., -50., 10., 10., -60.] # Simple model # theta = [.959, -46.271, 9.015, 8.531, -57.604] # skapar alla lanes, fences, roads, speed och trajectory for alla bilar for lane in lanes: r = r + theta[0] * lane.gaussian() for fence in fences: # increase the negative reward for the fences so that the cars dont go outside of the road #r = r+theta[1]*fence.gaussian()*1000000 r = r + theta[1] * fence.gaussian() if roads == None: pass else: for road in roads: r = r + theta[2] * road.gaussian(10.) if speed is not None: r = r + speed_import * theta[3] * feature.speed(speed) try: #quick fix, if there is just 1 car it will not be a list for traj in trajs: r = r + theta[4] * traj.gaussian() except: r = r + theta[4] * trajs.gaussian() return r
def simple_reward(self, trajs=None, lanes=None, roads=None, fences=None, speed=1., speed_import=1.): if lanes is None: lanes = self.lanes if roads is None: roads = self.roads if fences is None: fences = self.fences if trajs is None: trajs = [c.linear for c in self.cars] elif isinstance(trajs, car.Car): trajs = [c.linear for c in self.cars if c != trajs] r = 0.1 * feature.control() """ # What is theta? First one is importance of staying in lanes, second is staying on the road entirely (not violating the outer fence) third is staying on the road also? fourth is maintaining desired speed fifth is ...? """ theta = [1., -50., 10., 10., -60.] # Simple model # theta = [.959, -46.271, 9.015, 8.531, -57.604] for lane in lanes: r = r + theta[0] * lane.gaussian() for fence in fences: r = r + theta[1] * fence.gaussian() for road in roads: r = r + theta[2] * road.gaussian(10.) if speed is not None: r = r + speed_import * theta[3] * feature.speed(speed) for traj in trajs: r = r + theta[4] * traj.gaussian() return r
import lane dyn = dynamics.CarDynamics(0.1) vis = Visualizer(dyn.dt) vis.lanes.append(lane.StraightLane([0., -1.], [0., 1.], 0.13)) vis.lanes.append(vis.lanes[0].shifted(1)) vis.lanes.append(vis.lanes[0].shifted(-1)) vis.cars.append(car.UserControlledCar(dyn, [0., 0., math.pi / 2., .1])) vis.cars.append( car.SimpleOptimizerCar(dyn, [0., 0.5, math.pi / 2., 0.], color='red')) r = -60. * vis.cars[0].linear.gaussian() r = r + vis.lanes[0].gaussian() r = r + vis.lanes[1].gaussian() r = r + vis.lanes[2].gaussian() r = r - 30. * vis.lanes[1].shifted(1).gaussian() r = r - 30. * vis.lanes[2].shifted(-1).gaussian() r = r + 30. * feature.speed(0.5) r = r + 10. * vis.lanes[0].gaussian(10.) r = r + .1 * feature.control() vis.cars[1].reward = r vis.main_car = vis.cars[0] vis.paused = True vis.set_heat(r) #vis.set_heat(vis.lanes[0].gaussian()+vis.lanes[1].gaussian()+vis.lanes[2].gaussian()) #vis.set_heat(-vis.cars[1].traj.gaussian()+vis.lanes[0].gaussian()+vis.lanes[1].gaussian()+vis.lanes[2].gaussian()) vis.run() if __name__ == '__main__' and len(sys.argv) == 1: import world as wrld import car world = wrld.world2() vis = Visualizer(0.1, name='replay')
the_car = None for c in the_world.cars: if isinstance(c, car.UserControlledCar): the_car = c T = the_car.traj.T train = [] for fname in files: with open(fname) as f: us, xs = pickle.load(f) for t in range(T, len(xs[0]) - T, T): point = { 'x0': [xseq[t - 1] for xseq in xs], 'u': [useq[t:t + T] for useq in us] } train.append(point) theta = utils.vector(5) theta.set_value(np.array([1., -50., 10., 10., -60.])) r = 0.1 * feature.control() #features, thetas are weights for lane in the_world.lanes: r = r + theta[0] * lane.gaussian() for fence in the_world.fences: r = r + theta[1] * lane.gaussian() for road in the_world.roads: r = r + theta[2] * road.gaussian(10.) r = r + theta[3] * feature.speed(1.) for car in the_world.cars: if car != the_car: r = r + theta[4] * car.traj.gaussian() run_irl(the_world, the_car, r, theta, train)
if __name__ == '__main__' and False: import lane dyn = dynamics.CarDynamics(0.1) vis = Visualizer(dyn.dt) vis.lanes.append(lane.StraightLane([0., -1.], [0., 1.], 0.13)) vis.lanes.append(vis.lanes[0].shifted(1)) vis.lanes.append(vis.lanes[0].shifted(-1)) vis.cars.append(car.UserControlledCar(dyn, [0., 0., math.pi/2., .1])) vis.cars.append(car.SimpleOptimizerCar(dyn, [0., 0.5, math.pi/2., 0.], color='red')) r = -60.*vis.cars[0].linear.gaussian() r = r + vis.lanes[0].gaussian() r = r + vis.lanes[1].gaussian() r = r + vis.lanes[2].gaussian() r = r - 30.*vis.lanes[1].shifted(1).gaussian() r = r - 30.*vis.lanes[2].shifted(-1).gaussian() r = r + 30.*feature.speed(0.5) r = r + 10.*vis.lanes[0].gaussian(10.) r = r + .1*feature.control() vis.cars[1].reward = r vis.main_car = vis.cars[0] vis.paused = True vis.set_heat(r) #vis.set_heat(vis.lanes[0].gaussian()+vis.lanes[1].gaussian()+vis.lanes[2].gaussian()) #vis.set_heat(-vis.cars[1].traj.gaussian()+vis.lanes[0].gaussian()+vis.lanes[1].gaussian()+vis.lanes[2].gaussian()) vis.run() if __name__ == '__main__' and len(sys.argv)==1: import world as wrld import car world = wrld.world2() vis = Visualizer(0.1, name='replay')
t = idx_u*self.step_per_u+idx r_list.append(reward(t, self.x[t], self.u[idx_u])) #r = [reward(t, self.x[t], self.u[t]) for t in range(self.T)] return sum(r_list) """ g = [utils.grad(r[t], self.x[t]) for t in range(self.T)] for t in reversed(range(self.T-1)): g[t] = g[t]+tt.dot(g[t+1], utils.jacobian(self.x[t+1], self.x[t])) for t in range(self.T): g[t] = tt.dot(g[t], utils.jacobian(self.x[t], self.u[t]))+utils.grad(r[t], self.u[t], constants=[self.x[t]]) return sum(r), {self.u[t]: g[t] for t in range(self.T)} """ if __name__ == '__main__': from dynamics import CarDynamics import math dyn = CarDynamics(0.1) traj = Trajectory(5, dyn) l = lane.StraightLane([0., -1.], [0., 1.], .1) reward = feature.speed()+l.feature()#+feature.speed() r = traj.reward(reward) #traj.x0.value = np.asarray([0., 0., math.pi/2, 1.]) traj.x0.set_value([0.1, 0., math.pi/2, 1.]) optimizer = utils.Maximizer(r, traj.u) import time t = time.time() for i in range(1): optimizer.maximize(bounds=[(-1., 1.), (-2, 2.)]) print (time.time()-t)/1. print [u.get_value() for u in traj.u]
return f def reward(self, reward): r = [reward(t, self.x[t], self.u[t]) for t in range(self.T)] return sum(r) """ g = [utils.grad(r[t], self.x[t]) for t in range(self.T)] for t in reversed(range(self.T-1)): g[t] = g[t]+tt.dot(g[t+1], utils.jacobian(self.x[t+1], self.x[t])) for t in range(self.T): g[t] = tt.dot(g[t], utils.jacobian(self.x[t], self.u[t]))+utils.grad(r[t], self.u[t], constants=[self.x[t]]) return sum(r), {self.u[t]: g[t] for t in range(self.T)} """ if __name__ == '__main__': from dynamics import CarDynamics import math dyn = CarDynamics(0.1) traj = Trajectory(5, dyn) l = lane.StraightLane([0., -1.], [0., 1.], .1) reward = feature.speed()+l.feature()#+feature.speed() r = traj.reward(reward) #traj.x0.value = np.asarray([0., 0., math.pi/2, 1.]) traj.x0.set_value([0.1, 0., math.pi/2, 1.]) optimizer = utils.Maximizer(r, traj.u) import time t = time.time() for i in range(1): optimizer.maximize(bounds=[(-1., 1.), (-2, 2.)]) print (time.time()-t)/1. print [u.get_value() for u in traj.u]
# theta = [1., -50., 10., 10., -60.] # Simple model ======= # theta = [1., -50., 10., 10., -60., 10.] # Simple model # theta = [2.05026991,-50.,9.99045658,0.14135938,-60.] # Learned model # theta = [ 5.97469800e+00, -40.0789372, 10.0000000, .0168410493, -60.0000000] theta = [-118.675528, -49.9917950, 10.0000000, -.0158836823, -604.318363] # theta = [2.05026991,-50.,9.99045658,5,-60.] for lane in lanes: r = r+theta[0]*lane.gaussian() for fence in fences: r = r+theta[1]*fence.gaussian() for road in roads: r = r+theta[2]*road.gaussian(10.) if speed is not None: r = r+speed_import*theta[3]*feature.speed(speed) for traj in trajs: r = r+theta[4]*traj.gaussian() return r def playground(): dyn = dynamics.CarDynamics(0.1) world = World() clane = lane.StraightLane([0., -1.], [0., 1.], 0.17) world.lanes += [clane, clane.shifted(1), clane.shifted(-1)] world.roads += [clane] world.fences += [clane.shifted(2), clane.shifted(-2)] #world.cars.append(car.UserControlledCar(dyn, [0., 0., math.pi/2., 0.], color='orange')) world.cars.append(car.UserControlledCar(dyn, [-0.17, -0.17, math.pi/2., 0.], color='white')) return world
else: the_car = None for c in the_world.cars: if isinstance(c, car.UserControlledCar): the_car = c T = the_car.traj.T train = [] for fname in files: with open(fname) as f: us, xs = pickle.load(f) for t in range(T, len(xs[0])-T, T): point = { 'x0': [xseq[t-1] for xseq in xs], 'u': [useq[t:t+T] for useq in us] } train.append(point) theta = utils.vector(5) theta.set_value(np.array([1., -50., 10., 10., -60.])) r = 0.1*feature.control() for lane in the_world.lanes: r = r + theta[0]*lane.gaussian() for fence in the_world.fences: r = r + theta[1]*lane.gaussian() for road in the_world.roads: r = r + theta[2]*road.gaussian(10.) r = r + theta[3]*feature.speed(1.) for car in the_world.cars: if car!=the_car: r = r + theta[4]*car.traj.gaussian() run_irl(the_world, the_car, r, theta, train)