def test_import_box():
s = Simulator(mode='headless')
scene = StadiumScene()
s.import_scene(scene)
print(s.objects)
# wall = [pos, dim]
wall = [[[0, 7, 1.01], [10, 0.2, 1]], [[0, -7, 1.01], [6.89, 0.1, 1]],
[[7, -1.5, 1.01], [0.1, 5.5, 1]], [[-7, -1, 1.01], [0.1, 6, 1]],
[[-8.55, 5, 1.01], [1.44, 0.1, 1]], [[8.55, 4, 1.01], [1.44, 0.1, 1]]]
obstacles = [[[-0.5, 2, 1.01], [3.5, 0.1, 1]], [[4.5, -1, 1.01], [1.5, 0.1, 1]],
[[-4, -2, 1.01], [0.1, 2, 1]], [[2.5, -4, 1.01], [1.5, 0.1, 1]]]
for i in range(len(wall)):
curr = wall[i]
obj = BoxShape(curr[0], curr[1])
s.import_object(obj)
for i in range(len(obstacles)):
curr = obstacles[i]
obj = BoxShape(curr[0], curr[1])
s.import_object(obj)
config = parse_config(os.path.join(gibson2.root_path, '../test/test.yaml'))
turtlebot1 = Turtlebot(config)
turtlebot2 = Turtlebot(config)
s.import_robot(turtlebot1)
s.import_robot(turtlebot2)
turtlebot1.set_position([6., -6., 0.])
turtlebot2.set_position([-3., 4., 0.])
for i in range(100):
s.step()
s.disconnect()
def test_turtlebot_position():
s = Simulator(mode='headless')
scene = StadiumScene()
s.import_scene(scene)
turtlebot = Turtlebot(config)
s.import_robot(turtlebot)
turtlebot.set_position([0, 0, 5])
nbody = p.getNumBodies()
pos = turtlebot.get_position()
s.disconnect()
assert nbody == 5
assert np.allclose(pos, np.array([0, 0, 5]))
def test_import_building_viewing():
s = Simulator(mode='headless')
scene = BuildingScene('Ohoopee')
s.import_scene(scene)
assert s.objects == list(range(2))
turtlebot1 = Turtlebot(config)
turtlebot2 = Turtlebot(config)
turtlebot3 = Turtlebot(config)
s.import_robot(turtlebot1)
s.import_robot(turtlebot2)
s.import_robot(turtlebot3)
turtlebot1.set_position([0.5, 0, 0.5])
turtlebot2.set_position([0, 0, 0.5])
turtlebot3.set_position([-0.5, 0, 0.5])
for i in range(10):
s.step()
#turtlebot1.apply_action(np.random.randint(4))
#turtlebot2.apply_action(np.random.randint(4))
#turtlebot3.apply_action(np.random.randint(4))
s.disconnect()
def show_action_sensor_space():
s = Simulator(mode='headless')
scene = StadiumScene()
s.import_scene(scene)
turtlebot = Turtlebot(config)
s.import_robot(turtlebot)
turtlebot.set_position([0, 1, 0.5])
ant = Ant(config)
s.import_robot(ant)
ant.set_position([0, 2, 0.5])
h = Humanoid(config)
s.import_robot(h)
h.set_position([0, 3, 2])
jr = JR2(config)
s.import_robot(jr)
jr.set_position([0, 4, 0.5])
jr2 = JR2_Kinova(config)
s.import_robot(jr2)
jr2.set_position([0, 5, 0.5])
husky = Husky(config)
s.import_robot(husky)
husky.set_position([0, 6, 0.5])
quad = Quadrotor(config)
s.import_robot(quad)
quad.set_position([0, 7, 0.5])
for robot in s.robots:
print(type(robot), len(robot.ordered_joints), robot.calc_state().shape)
for i in range(100):
s.step()
s.disconnect()
def test_multiagent():
s = Simulator(mode='headless')
scene = StadiumScene()
s.import_scene(scene)
turtlebot1 = Turtlebot(config)
turtlebot2 = Turtlebot(config)
turtlebot3 = Turtlebot(config)
s.import_robot(turtlebot1)
s.import_robot(turtlebot2)
s.import_robot(turtlebot3)
turtlebot1.set_position([1, 0, 0.5])
turtlebot2.set_position([0, 0, 0.5])
turtlebot3.set_position([-1, 0, 0.5])
nbody = p.getNumBodies()
for i in range(100):
#turtlebot1.apply_action(1)
#turtlebot2.apply_action(1)
#turtlebot3.apply_action(1)
s.step()
s.disconnect()
assert nbody == 7
def rollout(traj, headless=False):
SCENE, traj_id, instr_id, instr, starting_coords = traj
# instr = "Keep going and don't stop"
# starting_coords = [0, 0, 0]
seq = tok.encode_sentence(instr)
tokens = tok.split_sentence(instr)
seq_lengths = [np.argmax(seq == padding_idx, axis=0)]
seq = torch.from_numpy(np.expand_dims(seq, 0)).cuda()
# seq_lengths[seq_lengths == 0] = seq.shape[1] # Full length
ctx, h_t, c_t, ctx_mask = encoder(seq, seq_lengths)
question = h_t
pre_feat = torch.zeros(batch_size, opts.img_feat_input_dim).cuda()
pre_ctx_attend = torch.zeros(batch_size, opts.rnn_hidden_size).cuda()
# Gibson stuff
# 72 fov for 600, 60 for 480
# mode = gui for debug, headless for run
s = Simulator(mode='gui', resolution=640, fov=75, panorama=True)
scene = BuildingScene(SCENE)
# scene = StadiumScene()
ids = s.import_scene(scene)
robot = Turtlebot(config)
ped_id = s.import_robot(robot)
heading_feat_tensor = torch.Tensor(heading_elevation_feat()).view([im_per_ob, 128]).cuda()
s.step()
robot.set_position(starting_coords)
def apply_action(bot: robot, action_idx: int, depth_ok: list, headless=False) -> bool:
print(action_idx)
# action_idx is expected to be 0-13, TODO: make nicer...
if action_idx == 0 or action_idx > 12 or not depth_ok[action_idx - 1]:
print("STOP")
return True
action_idx -= 1
#if action_idx < 3 or (action_idx < 12 and action_idx > 9):
bot.turn_right(0.5235988 * action_idx)
s.step()
if(not headless):
time.sleep(0.2)
bot.move_forward(0.5)
return False
# else:
# if action_idx < 7:
# bot.turn_left(1.57)
# else:
# bot.turn_right(1.57)
bot_is_running = True
while bot_is_running:
s.step()
gib_out = s.renderer.render_robot_cameras(modes=('rgb', '3d'))
rgb = gib_out[::2]
depth = np.array(gib_out[1::2])
processed_rgb = list(map(transform_img, rgb))
batch_obs = np.concatenate(processed_rgb)
imgnet_input = torch.Tensor(batch_obs).cuda()
imgnet_output = torch.zeros([im_per_ob, 2048]).cuda()
# depth processing and filtering
# depth: [36, ]
depth *= depth
depth = depth[:, :, :, :3].sum(axis=3)
depth = np.sqrt(depth)
# filter out 0 distances that are presumably from infinity dist
depth[depth < 0.0001] = 10
# TODO: generalize to non-horizontal moves
depth_ok = depth[12:24, 200:440, 160:480].min(axis=2).min(axis=1)
fig=plt.figure(figsize=(8, 2))
for n, i in enumerate([0, 3, 6, 9]):
fig.add_subplot(1, 4, n + 1)
plt.imshow(depth[12 + i])
plt.show()
# depth_ok *= depth_ok > 1
print(depth_ok)
depth_ok = depth_ok > 0.8
print(depth_ok)
# Each observation has 36 inputs
# We pass rgb images through frozen embedder
for i in range(im_per_ob // B_S):
def hook_fn(m, last_input, o):
imgnet_output[i*B_S:(i+1)*B_S, :] = \
o.detach().squeeze(2).squeeze(2)
imgnet_input[B_S * i : B_S * (i + 1)]
# imgnet_output[B_S * i : B_S * (i + 1)] = resnet(minibatch).detach()
imgnet_output = torch.cat([imgnet_output, heading_feat_tensor], 1)
pano_img_feat = imgnet_output.view([1, im_per_ob, 2176])
navigable_feat = torch.zeros([1, 16, 2176]).cuda()
navigable_feat[0, 1:13] = imgnet_output[12:24] * torch.Tensor(depth_ok).cuda().view(12, 1)
# TODO: make nicer as stated above
navigable_index = [list(map(int, depth_ok))]
print(navigable_index)
# NB: depth_ok replaces navigable_index
h_t, c_t, pre_ctx_attend, img_attn, ctx_attn, logit, value, navigable_mask = model(
pano_img_feat, navigable_feat, pre_feat, question, h_t, c_t, ctx,
pre_ctx_attend, navigable_index, ctx_mask)
print("ATTN")
print(ctx_attn[0])
print(img_attn[0])
plt.bar(range(len(tokens)), ctx_attn.detach().cpu()[0][:len(tokens)])
plt.xticks(range(len(tokens)), tokens)
plt.show()
plt.bar(range(16), img_attn.detach().cpu()[0])
plt.show()
print("NMASK")
print(navigable_mask)
logit.data.masked_fill_((navigable_mask == 0).data, -float('inf'))
m = torch.Tensor([[False] + list(map(lambda b: not b, navigable_index[0])) + [False, False, False]], dtype=bool).cuda()
logit.data.masked_fill_(m, -float('inf'))
action = _select_action(logit, [False])
ended = apply_action(robot, action[0], depth_ok)
bot_is_running = not ended or not headless
if not headless:
time.sleep(.3)