File size: 16,491 Bytes
15b8951 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 | import json
import math
import threading
import numpy as np
import pytest
from vlm.webconsole.demo1.navloop import (
parse_vla_json, build_nav_prompt, VlaError, NavLoopAgent,
MAX_STEP_METERS, parse_navila_action,
)
# --- parse_navila_action (NaVILA ngôn ngữ tự nhiên -> schema JSON) ---------- #
def test_navila_move_forward_cm():
s = parse_navila_action("The next action is move forward 25 cm.")
step = parse_vla_json(s)
assert step.action == "move_forward"
assert math.isclose(step.to_motion_cmd().value, 0.25)
def test_navila_turn_left_degrees():
step = parse_vla_json(parse_navila_action("turn left 15 degrees"))
assert step.action == "turn_left"
assert math.isclose(step.to_motion_cmd().value, math.radians(15))
def test_navila_turn_right_degrees():
step = parse_vla_json(parse_navila_action("The next action is turn right 30 degrees."))
assert step.action == "turn_right" and step.to_motion_cmd().value < 0
def test_navila_stop_is_finished():
step = parse_vla_json(parse_navila_action("stop, the task is completed."))
assert step.finished and step.to_motion_cmd() is None
def test_navila_unparsed_falls_back_to_stop():
# Câu KHÔNG có hướng -> DỪNG an toàn.
step = parse_vla_json(parse_navila_action("hmm not sure what to do"))
assert step.finished
def test_navila_loose_forward_without_number():
# NaVILA hay ra "move forward a certain distance" (thiếu số) -> loose gán bước mặc định.
step = parse_vla_json(parse_navila_action("The next action is move forward a certain distance."))
assert step.action == "move_forward" and not step.finished
def test_navila_loose_ambiguous_left_right_stops():
# "turn left or right" mơ hồ -> KHÔNG đoán bừa -> stop an toàn.
step = parse_vla_json(parse_navila_action("turn left or right by a specific degree"))
assert step.finished
# --- parse_vla_json -------------------------------------------------------- #
def test_parse_move_forward_cm():
s = parse_vla_json('{"action":"move_forward","value":75,"unit":"cm",'
'"reasoning":"đi tới","is_finished":false}')
assert s.action == "move_forward" and not s.finished
mc = s.to_motion_cmd()
assert mc.kind == "move" and math.isclose(mc.value, 0.75)
def test_parse_move_backward_negative():
s = parse_vla_json('{"action":"move_backward","value":0.3,"unit":"m"}')
assert math.isclose(s.to_motion_cmd().value, -0.3)
def test_parse_turn_left_degrees_positive():
s = parse_vla_json('{"action":"turn_left","value":30,"unit":"degrees"}')
assert math.isclose(s.to_motion_cmd().value, math.radians(30))
def test_parse_turn_right_negative():
s = parse_vla_json('{"action":"turn_right","value":45,"unit":"deg"}')
assert math.isclose(s.to_motion_cmd().value, -math.radians(45))
def test_parse_strips_markdown_fence():
s = parse_vla_json('```json\n{"action":"stop","is_finished":true,'
'"reasoning":"tới rồi"}\n```')
assert s.finished and s.to_motion_cmd() is None
def test_parse_is_finished_without_action_is_stop():
s = parse_vla_json('{"is_finished":true,"reasoning":"done"}')
assert s.finished
def test_parse_obstacles_list():
s = parse_vla_json('{"action":"stop","obstacles_detected":["chair","table"]}')
assert s.obstacles == ["chair", "table"]
def test_parse_clamps_huge_distance():
s = parse_vla_json('{"action":"move_forward","value":999,"unit":"m"}')
assert abs(s.to_motion_cmd().value) <= MAX_STEP_METERS
def test_parse_no_json_raises():
with pytest.raises(VlaError):
parse_vla_json("xin chào, không có json ở đây")
def test_parse_invalid_action_raises():
with pytest.raises(VlaError):
parse_vla_json('{"action":"teleport","value":5,"unit":"m"}')
def test_parse_bad_value_raises():
with pytest.raises(VlaError):
parse_vla_json('{"action":"move_forward","value":"abc","unit":"m"}')
# --- build_nav_prompt ------------------------------------------------------ #
def test_prompt_has_schema_and_goal():
p = build_nav_prompt("đi tới ghế", history=["move forward 0.50 m"],
obstacle_hint="chair")
assert "đi tới ghế" in p
assert "move_forward" in p and "is_finished" in p
assert "move forward 0.50 m" in p and "chair" in p
# --- NavLoopAgent loop ----------------------------------------------------- #
class FakeBrain:
"""Trả lần lượt các câu JSON đã định sẵn."""
def __init__(self, outputs):
self.outputs = list(outputs)
self.calls = 0
def decide(self, frame, prompt):
out = self.outputs[self.calls]
self.calls += 1
return out
class FakeMotion:
def __init__(self, error_on=None):
self.error_on = error_on # message lỗi để mô phỏng (vd vật cản)
self.ran = []
def run(self, cmd, cancel=None):
self.ran.append(cmd)
yield {"type": "step", "id": "motion", "status": "running", "title": "go"}
if self.error_on:
yield {"type": "step", "id": "motion", "status": "error"}
yield {"type": "error", "message": self.error_on}
return
yield {"type": "step", "id": "motion", "status": "done"}
yield {"type": "answer", "text": "Done.", "state": "UNKNOWN"}
class FakeFrames:
def get_latest_frame(self):
return object() # frame giả; FakeBrain không dùng tới
def _answers(events):
return [e for e in events if e.get("type") == "answer"]
def test_loop_moves_then_finishes(tmp_path):
brain = FakeBrain([
'{"action":"move_forward","value":50,"unit":"cm","reasoning":"tiến"}',
'{"action":"turn_left","value":30,"unit":"degrees","reasoning":"xoay"}',
'{"action":"stop","is_finished":true,"reasoning":"đã tới ghế"}',
])
motion = FakeMotion()
agent = NavLoopAgent(brain, FakeFrames(), motion, log_dir=str(tmp_path), settle_s=0)
events = list(agent.run("đi tới ghế"))
ans = _answers(events)
assert ans and ans[-1]["state"] == "YES"
assert ans[-1]["text"] == "đã tới ghế"
assert len(motion.ran) == 2 # 2 lệnh chuyển động trước stop
# motion's own "answer" must be swallowed -> chỉ 1 answer cuối cùng.
assert len(ans) == 1
def test_loop_writes_jsonl_log(tmp_path):
brain = FakeBrain([
'{"action":"move_forward","value":50,"unit":"cm","reasoning":"tiến"}',
'{"action":"stop","is_finished":true,"reasoning":"xong"}',
])
agent = NavLoopAgent(brain, FakeFrames(), FakeMotion(), log_dir=str(tmp_path), settle_s=0)
list(agent.run("đi"))
logs = list(tmp_path.glob("nav_*.jsonl"))
assert len(logs) == 1
lines = [json.loads(l) for l in logs[0].read_text(encoding="utf-8").splitlines()]
assert lines[0]["event"] == "start"
assert any(r.get("action") == "move_forward" for r in lines[1:])
def test_loop_obstacle_reobserves_not_terminate(tmp_path):
# Bước 1 đi -> vật cản (không kết thúc), bước 2 stop.
brain = FakeBrain([
'{"action":"move_forward","value":100,"unit":"cm","reasoning":"thử đi"}',
'{"action":"stop","is_finished":true,"reasoning":"né xong, dừng"}',
])
motion = FakeMotion(error_on="⛔ Vật cản phía trước — đã dừng an toàn.")
agent = NavLoopAgent(brain, FakeFrames(), motion, log_dir=str(tmp_path), settle_s=0)
events = list(agent.run("đi tới"))
ans = _answers(events)
assert ans[-1]["state"] == "YES" # vẫn chạy tới bước stop
assert brain.calls == 2 # đã quan sát lại sau vật cản
def test_loop_terminates_on_motion_timeout(tmp_path):
brain = FakeBrain([
'{"action":"move_forward","value":50,"unit":"cm","reasoning":"đi"}',
'{"action":"stop","is_finished":true,"reasoning":"không nên tới đây"}',
])
motion = FakeMotion(error_on="Quá thời gian chuyển động.")
agent = NavLoopAgent(brain, FakeFrames(), motion, log_dir=str(tmp_path), settle_s=0)
events = list(agent.run("đi"))
assert any(e.get("type") == "error" for e in events)
assert brain.calls == 1 # dừng ngay, không quan sát lại
def test_loop_respects_cancel(tmp_path):
brain = FakeBrain(['{"action":"move_forward","value":50,"unit":"cm"}'] * 5)
cancel = threading.Event()
cancel.set()
agent = NavLoopAgent(brain, FakeFrames(), FakeMotion(), log_dir=str(tmp_path), settle_s=0)
events = list(agent.run("đi", cancel=cancel))
assert any("dừng theo yêu cầu" in e.get("message", "") for e in events)
assert brain.calls == 0
def test_loop_max_steps_cap(tmp_path):
# Luôn move_forward, không bao giờ stop -> phải dừng ở max_steps.
brain = FakeBrain(['{"action":"move_forward","value":50,"unit":"cm"}'] * 10)
agent = NavLoopAgent(brain, FakeFrames(), FakeMotion(), max_steps=3,
log_dir=str(tmp_path), settle_s=0)
events = list(agent.run("đi"))
assert brain.calls == 3
assert _answers(events)[-1]["state"] == "UNKNOWN"
class _AnnHint:
raw_box = (1, 1, 2, 2)
label = "bottle"
def test_obstacle_hint_from_annotator(tmp_path):
# Hint lấy từ annotator (1 nguồn YOLO), KHÔNG detect riêng.
agent = NavLoopAgent(FakeBrain([]), FakeFrames(), FakeMotion(),
annotator=_AnnHint(), log_dir=str(tmp_path), settle_s=0)
assert agent._obstacle_hint(None) == "bottle"
# không có annotator -> None
agent2 = NavLoopAgent(FakeBrain([]), FakeFrames(), FakeMotion(),
log_dir=str(tmp_path), settle_s=0)
assert agent2._obstacle_hint(None) is None
def test_loop_bad_json_terminates(tmp_path):
brain = FakeBrain(["không phải json"])
agent = NavLoopAgent(brain, FakeFrames(), FakeMotion(), log_dir=str(tmp_path), settle_s=0)
events = list(agent.run("đi"))
assert any("JSON VLM hỏng" in e.get("message", "") for e in events)
# --- YOLO override điểm dừng ----------------------------------------------- #
class FakeYolo:
"""detect(wanted) trả box 'bottle' với y2 lấy lần lượt từ y2_list (mô phỏng
robot tiến lại gần -> đáy hộp tụt dần về mép dưới ảnh). Call obstacle-hint
(wanted=None) trả box cố định, KHÔNG tiêu thụ y2_list."""
names = ["bottle", "chair"]
def __init__(self, y2_list):
self.y2 = list(y2_list)
self.n = 0
def detect(self, frame, wanted=None):
if not wanted:
return [(100, 100, 200, 250, "bottle", 0.9)]
y2 = self.y2[min(self.n, len(self.y2) - 1)]
self.n += 1
return [(100, 100, 200, y2, "bottle", 0.9)]
class ShapedFrames:
def get_latest_frame(self):
return np.zeros((720, 1280, 3), dtype=np.uint8) # khớp camera thật 1280x720
def test_no_override_when_target_at_bottom(tmp_path):
brain = FakeBrain(['{"action":"stop","is_finished":true,"reasoning":"tới"}'])
motion = FakeMotion()
yolo = FakeYolo([460]) # gap 20 <= 50 -> chấp nhận dừng
agent = NavLoopAgent(brain, ShapedFrames(), motion, detector=yolo,
log_dir=str(tmp_path), settle_s=0, stop_bottom_px=50)
events = list(agent.run("go to the bottle"))
assert not any("YOLO override" in e.get("text", "") for e in events if e.get("type") == "token")
assert len(motion.ran) == 0
assert _answers(events)[-1]["state"] == "YES"
# --- servo hình học (ưu tiên: stop > center > forward) --------------------- #
def _servo_agent(tmp_path):
return NavLoopAgent(FakeBrain([]), ShapedFrames(), FakeMotion(),
log_dir=str(tmp_path), settle_s=0, stop_bottom_px=20)
class SeqAnn:
"""annotator giả: trả box lần lượt theo list (mô phỏng robot tiến lại gần)."""
label = "bottle"
def __init__(self, boxes):
self.boxes = boxes
self.i = 0
def target_box(self):
b = self.boxes[min(self.i, len(self.boxes) - 1)]
self.i += 1
return b
def frame_height(self):
return 720
def frame_width(self):
return 1280
class ServoMotion:
"""Ghi lại các (vx, wz) được bơm + số lần stop (cho visual servo liên tục)."""
def __init__(self):
self.pubs = []
self.stops = 0
def _publish(self, vx, wz):
self.pubs.append((round(vx, 3), round(wz, 3)))
def _stop(self):
self.stops += 1
def front_distance(self):
return None
def test_visual_servo_turn_then_forward_then_stop(tmp_path):
# run() tiêu thụ box[0]; servo: lệch -> XOAY (vx=0); giữa -> TIẾN (wz=0); đáy chạm -> DỪNG.
ann = SeqAnn([(1120, 100, 1220, 300), # run() đọc trước
(1120, 100, 1220, 300), # servo: lệch ~37° (>30) -> CHỈ xoay (vx=0)
(600, 100, 680, 300), # servo: giữa (lệch 0) -> tiến thẳng (wz=0)
(600, 100, 680, 710)]) # servo: đáy 10px <= 20 -> dừng
motion = ServoMotion()
ag = NavLoopAgent(FakeBrain([]), ShapedFrames(), motion, annotator=ann,
log_dir=str(tmp_path), settle_s=0, stop_bottom_px=20, control="servo")
ag.servo_hz = 1000.0; ag.servo_pulse_s = 0; ag.servo_refresh_s = 0; ag.stop_settle_s = 0; ag.final_push_m = 0 # chạy nhanh cho test
events = list(ag.run("go to the bottle"))
assert _answers(events)[-1]["state"] == "YES"
assert any(vx == 0.0 and wz != 0.0 for vx, wz in motion.pubs) # P1 xoay căn giữa
assert any(vx > 0.0 and wz == 0.0 for vx, wz in motion.pubs) # P2 tiến thẳng
assert motion.stops >= 1
def test_visual_servo_stops_immediately_at_bottom(tmp_path):
# Đáy đã sát mép ngay từ đầu -> servo DỪNG, không bơm vx tiến.
ann = SeqAnn([(600, 100, 680, 712), (600, 100, 680, 712)])
motion = ServoMotion()
ag = NavLoopAgent(FakeBrain([]), ShapedFrames(), motion, annotator=ann,
log_dir=str(tmp_path), settle_s=0, stop_bottom_px=20, control="servo")
ag.servo_hz = 1000.0; ag.servo_pulse_s = 0; ag.servo_refresh_s = 0; ag.stop_settle_s = 0; ag.final_push_m = 0
events = list(ag.run("go to the bottle"))
assert _answers(events)[-1]["state"] == "YES"
assert not any(vx > 0.0 for vx, wz in motion.pubs) # không tiến (đã sát đáy)
assert motion.stops >= 1
def test_state_text_injects_geometry(tmp_path):
ag = _servo_agent(tmp_path)
s = ag._state_text((1000, 100, 1100, 300), 1280, 720, 1.5)
assert "lệch" in s and "phải" in s
assert "đáy cách mép dưới 420px" in s and "1.50m" in s
def test_vlm_turn_capped_to_max(tmp_path):
# VLM xuất turn 90° -> code chặn xuống ≤ max_turn_deg (10°).
brain = FakeBrain(['{"action":"turn_left","value":90,"unit":"degrees"}'])
motion = FakeMotion()
ag = NavLoopAgent(brain, FakeFrames(), motion, max_steps=1,
log_dir=str(tmp_path), settle_s=0, control="vlm")
list(ag.run("scan"))
assert len(motion.ran) == 1
assert abs(motion.ran[0].value) <= math.radians(10) + 1e-6
def test_search_accumulator_stops_after_full_scan(tmp_path):
# Quét tìm xoay mãi không thấy -> đủ ngưỡng góc thì DỪNG (chống xoay vô tận).
brain = FakeBrain(['{"action":"turn_left","value":10,"unit":"degrees"}'] * 10)
motion = FakeMotion()
ag = NavLoopAgent(brain, FakeFrames(), motion, log_dir=str(tmp_path),
settle_s=0, control="vlm")
ag.search_max_deg = 25 # 10+10+10=30 >= 25 -> dừng ở bước 3
events = list(ag.run("go to the bottle"))
assert any("Không tìm thấy mục tiêu" in e.get("message", "") for e in events)
assert len(motion.ran) <= 3
def test_vlm_trusts_stop_pure_vlm(tmp_path):
# THUẦN VLM: VLM báo stop -> tin ngay (không còn code override).
brain = FakeBrain(['{"action":"stop","is_finished":true,"reasoning":"tới"}'])
agent = NavLoopAgent(brain, FakeFrames(), FakeMotion(), log_dir=str(tmp_path),
settle_s=0, control="vlm")
events = list(agent.run("go to the bottle"))
assert _answers(events)[-1]["state"] == "YES"
|