-
Notifications
You must be signed in to change notification settings - Fork 6
/
embodied_vln.py
221 lines (170 loc) · 6.94 KB
/
embodied_vln.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
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
import os
import cv2
import anthropic
import numpy as np
from PIL import Image
from vln.agent import AirsimAgent
import base64
from openai import OpenAI
from prompts.prompt2 import build_prompt
from utils import encode_image, LM_VLN
def parse_llm_action(llm_output: str):
command_str = llm_output.split(":")[-1]
command_str = command_str.strip(" ")
command_str = command_str.lower()
act_enum = -1
if 'stop' in command_str:
return 0
elif 'forth' in command_str:
return 1
elif 'left' in command_str:
return 2
elif 'right' in command_str:
return 3
elif 'up' in command_str:
return 4
elif 'down' in command_str:
return 5
else:
return -1
class VLN_evaluator:
def __init__(self, root_dir, eval_model, api_key):
self.root_dir = root_dir
self.eval_model = eval_model
self.agent = AirsimAgent(None, None, None)
self.navi_tasks = self.load_navi_task()
self.client = LM_VLN(eval_model, api_key)
def load_navi_task(self):
navi_data = []
task_file = os.path.join(self.root_dir, 'start_loc.txt')
gt_traj_dir = os.path.join(self.root_dir, 'label')
if not os.path.isfile(task_file):
raise ValueError(f"Task file not found in {task_file}")
with open(task_file, 'r') as f:
task_data = f.readlines()
traj_files = os.listdir(gt_traj_dir)
traj_files = sorted(traj_files, key=lambda x: int(x.split('.')[0]))
assert len(task_data) == len(traj_files)
for i in range(len(task_data)):
task_line = task_data[i]
traj_file = os.path.join(self.root_dir, 'label', traj_files[i])
init_pos, init_rot, task_desc = self.parse_task_line(task_line)
gt_traj = self.parse_traj_file(traj_file)
target_pos = init_pos + gt_traj[len(gt_traj)-1]
gt_traj_len = 0.0
last_pos = np.zeros(3)
for j in range(len(gt_traj)):
step_len = np.linalg.norm(gt_traj[j] - last_pos)
gt_traj_len += step_len
last_pos = gt_traj[j]
navi_data.append({
"start_pos": init_pos,
"target_pos": target_pos,
"start_rot": init_rot,
"gt_traj": gt_traj,
"gt_traj_len": gt_traj_len,
"task_desc": task_desc
})
return navi_data
def parse_task_line(self, task_line: str):
task_line = task_line.strip('\n')
items = task_line.split(';')
pos_corp = items[0].strip(' ')
rot_corp = items[1].strip(' ')
desc = items[2].strip(' ')
pos_str_items = pos_corp.split(' ')[1:]
rot_str_items = rot_corp.split(':')[1].split(', ')
for i in range(len(pos_str_items)):
pos_str_items[i] = pos_str_items[i].strip(',')
for i in range(len(rot_str_items)):
rot_str_items[i] = rot_str_items[i].strip(' ')
pos = list(map(float, pos_str_items))
rot = list(map(float, rot_str_items))
pos = np.array(pos) / 100 # cm to m
rot = np.array(rot)
return pos, rot, desc
def parse_traj_file(self, traj_file: str):
if not os.path.isfile(traj_file):
raise ValueError(f"Trajectory file is not found in {traj_file}")
with open(traj_file, 'r') as f:
traj_lines = f.readlines()
traj = []
traj_lines = traj_lines[1:]
for i in range(len(traj_lines)):
traj_line = traj_lines[i].strip('\n')
pos_str_items = traj_line.split(',')[1:]
pos = list(map(float, pos_str_items))
traj.append(pos)
return np.array(traj)
def evaluation(self):
navi_data = self.navi_tasks
SR_count = 0.0
SPL = 0.0
traj_len = 0.0
ne_count = 0.0
SR_short_idx = []
SR_long_idx = []
for idx, navi_task in enumerate(navi_data):
if idx > 10:
break
traj_len = 0.0
start_pos = navi_task["start_pos"]
start_rot = navi_task["start_rot"]
gt_traj = navi_task["gt_traj"]
target_pos = navi_task["target_pos"]
gt_traj_len = navi_task["gt_traj_len"]
task_desc = navi_task["task_desc"]
start_pos[2] = -start_pos[2] # unreal coords to airsim coords
start_pose = np.concatenate((start_pos, start_rot))
# print(f"start pose: {start_pose}")
self.agent.setVehiclePose(start_pose)
# time.sleep(1)
# self.agent.client.moveToPositionAsync(float(start_pose[0]), float(start_pose[1]), float(start_pose[2]), 1).join()
pos, rot = self.agent.get_current_state()
print(f"pos: {pos}, rot: {rot}")
messages = []
step_size = 0
while step_size < 30:
answer = self.client.query(self.agent, messages, task_desc)
# print(answer)
act = parse_llm_action(answer)
print("action: ", act)
if act == 0:
break
self.agent.makeAction(act)
cur_pos, cur_rot = self.agent.get_current_state()
if act in [1, 4, 5]:
traj_len += 10.0
step_size += 1
dist = np.linalg.norm(cur_pos - target_pos)
print(f"Task idx: {idx}, current step size: {step_size}, current dist: {dist}")
if dist < 20:
break
elif dist > 300:
break
print(f"Max step size reached or target reached. step size: {step_size}")
final_pos, final_rot = self.agent.get_current_state()
dist = np.linalg.norm(final_pos - target_pos)
if dist < 20:
if gt_traj_len > 100:
SR_long_idx.append(idx)
else:
SR_short_idx.append(idx)
SR_count += 1
SPL_count = gt_traj_len / max(gt_traj_len, traj_len)
SPL += SPL_count
ne_count += dist
print(f"####### SR count: {SR_count}, SPL: {SPL}, NE: {ne_count}")
# time.sleep(10)
SR = SR_count / len(navi_data)
NE = ne_count / len(navi_data)
print(f"SR: {SR}, SPL: {SPL}, NE: {NE}")
print(SR_short_idx)
print(SR_long_idx)
if __name__ == "__main__":
model = "xxxxx" # LM models, for example: "claude-3-haiku-20240307", "gpt-4o"
api_key = "xxxxxxxxx" # Fill in API key
vln_eval = VLN_evaluator("dataset/vln", model, api_key)
navi_data = vln_eval.navi_tasks
vln_eval.evaluation()
# print(navi_data[0])