-
Notifications
You must be signed in to change notification settings - Fork 0
/
task_mh2.py
109 lines (84 loc) · 4.46 KB
/
task_mh2.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
import numpy as np
from physics_sim import PhysicsSim
class Task_mh():
"""Task (environment) that defines the goal and provides feedback to the agent."""
def __init__(self, init_pose=None, init_velocities=None,
init_angle_velocities=None, runtime=5., target_pos=None):
"""Initialize a Task object.
Params
======
init_pose: initial position of the quadcopter in (x,y,z) dimensions and the Euler angles
init_velocities: initial velocity of the quadcopter in (x,y,z) dimensions
init_angle_velocities: initial radians/second for each of the three Euler angles
runtime: time limit for each episode
target_pos: target/goal (x,y,z) position for the agent
"""
# Simulation
self.sim = PhysicsSim(init_pose, init_velocities, init_angle_velocities, runtime)
self.action_repeat = 3
self.state_size = self.action_repeat * 6
self.action_low = 0
self.action_high = 900
self.action_size = 4
self.take_off = np.array([0., 0., 20.])
# Goal
self.target_pos = target_pos if target_pos is not None else np.array([0., 0., 10.])
def get_reward(self):
"""Uses current pose of sim to return reward."""
# ----- Approach #1 -----
# reward = 1.-.3*(abs(self.sim.pose[:3] - self.target_pos)).sum()
# ----- Approach #2 -----
# the reward function should be normalized between -1 and 1 (except for colisions) in order to the NN better learn the gradients. The hyperbolic tangent function np.tanh is used for this purpose. But that didn't produce anything efficient.
# reward = np.tanh(1-0.03*(abs(self.sim.pose[:3] - self.target_pos)).sum())
# ----- Approach #3 -----
# reward = 0
# approach_x = abs(self.sim.pose[0]-self.target_pos[0])**2
# approach_y = abs(self.sim.pose[1]-self.target_pos[1])**2
# approach_z = abs(self.sim.pose[2]-self.target_pos[2])**2
# dist_quadtotar = np.sqrt((self.sim.pose[0]-self.target_pos[0])**2 + (self.sim.pose[1]-self.target_pos[1])**2 + (self.sim.pose[2]-self.target_pos[2])**2)
# positive reward when take-off.
# if self.sim.pose[2] > self.take_off[2]:
# reward += 1.1
# positive reward when higher z than target_pos.
# elif self.sim.pose[2] > self.target_pos[2]:
# reward += 1.3
# # positive reward when close to target_pos.
# if dist_quadtotar < 10:
# reward += 70 # 100 500
# # positive reward when close to target_pos.
# if approach_x < 10:
# reward += 50 # 100 500
# # positive reward when close to target_pos.
# elif approach_y < 10:
# reward += 50 # 100 500
# # positive reward when close to target_pos.
# elif approach_z < 10:
# reward += 50 # 100 500
# # negative reward if flying but not reaching the Z of target_pos.
# # elif self.target_pos[2] > self.sim.pose[2]:
# # reward += -1.3
# # negative reward if flying but not reaching the target_pos.
# elif self.target_pos[2] > self.sim.pose[2] and dist_quadtotar >= 10:
# reward += -1.3
# negative reward if done before the runtime finished and stop
# elif self.sim.done and self.sim.runtime > self.sim.time:
# reward += -2 # -1
# ----- Approach #4 -----
# reward = np.tanh(1-0.0046*(abs(self.sim.pose[:3] - self.target_pos)).sum())
reward = np.tanh(1-0.005*(abs(self.sim.pose[:3] - self.target_pos)).sum())
return reward
def step(self, rotor_speeds):
"""Uses action to obtain next state, reward, done."""
reward = 0
pose_all = []
for _ in range(self.action_repeat):
done = self.sim.next_timestep(rotor_speeds) # update the sim pose and velocities
reward += self.get_reward()
pose_all.append(self.sim.pose)
next_state = np.concatenate(pose_all)
return next_state, reward, done
def reset(self):
"""Reset the sim to start a new episode."""
self.sim.reset()
state = np.concatenate([self.sim.pose] * self.action_repeat)
return state