def start(self, fps, fast_step=False):
"""Start
Begin the continuous event loop for the simulation. This event loop
can be exited using the ctrl+c keyboard interrupt. Real-time
constraints are enforced. [Hz]
Arguments:
fps: The value of frames per second of the simulation.
fast_step: If True, the ODE fast step algorithm will be used.
This is faster and requires less memory but is less accurate.
(Default: False)
"""
paused = False
stopped = False
# Define the total time for the tooltip traversal
t_total = 20.0
# Define the simulation frame rate
t = 0.0 # [s]
dt = 1.0 / float(fps) # [s]
# Keep track of time overshoot in the case that simulation time must be
# increased in order to maintain real-time constraints
t_overshoot = 0.0
# Get the initial path position (center of gate7)
pos_start = self.env.get_body_pos('gate7') # [m]
# Get the first position of the PA10 at rest
pos_init = self.env.get_body_pos('tooltip') # [m]
# Calculate the new required joint angles of the PA10
#pa10_joint_angles = self.kinematics.calc_inverse_kinematics(pos_init, pos_start)
# TODO: Move the PA10 end-effector to the starting position along the path
# TODO: TEMP - Move the temporary end-effector pointer to the starting position
self.env.set_group_pos('pointer', pos_start)
# Generate long-term path from initial position
t_input = np.linspace(start=0.0, stop=1.0, num=t_total / dt)
t_input = np.reshape(t_input, (len(t_input), 1))
rnn_path = self.rnn.extrapolate(t_input, [pos_start], len(t_input) - 1)
# Add the initial condition point back onto the data
rnn_path = np.vstack((pos_start, rnn_path))
# Retrieve one set of standard gate position/orientation data
file_path = pathutils.list_data_files(constants.G_TRAINING_DATA_DIR)[0]
gate_data = datastore.retrieve(file_path)
gate_start_idx = constants.G_GATE_IDX
gate_end_idx = gate_start_idx + constants.G_NUM_GATE_INPUTS
# Reshape the gate positions data
gate_data = gate_data[0:1, gate_start_idx:gate_end_idx]
gate_data = np.tile(gate_data, (len(rnn_path), 1))
# Complete the rnn path data
rnn_path = np.hstack((t_input, gate_data, rnn_path))
# Save generated path for later examination
datastore.store(rnn_path, constants.G_RNN_STATIC_PATH_OUT)
# Define a variable to hold the final path (with real-time correction)
final_path = rnn_path[:-1].copy()
path_saved = False
# Detect all path segments between gates in the generated path
segments = pathutils._detect_segments(rnn_path)
path_idx = 0
x_path_offset = np.array([0.0, 0.0, 0.0]) # [m]
v_curr = np.array([0.0, 0.0, 0.0]) # [m/s]
a_max = constants.G_MAX_ACCEL # [m/s^2]
# Get the static table position
x_table = self.env.get_body_pos('table')
while not stopped:
t_start = time.time()
# If the last calculation took too long, catch up
dt_warped = dt + t_overshoot
self.env.set_dt(dt_warped)
# Determine if the viewer is stopped. Then we can quit
if self.viewer.is_dead:
break
# Pause the simulation if we are at the end
if path_idx == len(rnn_path) - 1 or paused:
self.env.step(paused=True, fast=fast_step)
# If we have really hit the end of the simulation, save/plot the path
if not paused and not path_saved:
# Save the final data to a file
datastore.store(final_path,
constants.G_RNN_DYNAMIC_PATH_OUT)
path_saved = True
continue
# Not a very elegant solution to pausing at the start, but it works
if t <= 1000.0:
self.env.step(paused=True, fast=fast_step)
t += dt_warped
continue
# Determine the current path segment
curr_segment_idx = 0
for segment_idx, segment_end in enumerate(segments):
if path_idx <= segment_end:
curr_segment_idx = segment_idx
break
x_curr = pathutils.get_path_tooltip_pos(rnn_path,
path_idx) + x_path_offset
x_next = pathutils.get_path_tooltip_pos(
rnn_path, path_idx + 1) + x_path_offset
# Get the expected gate position
x_gate_expected = pathutils.get_path_gate_pos(
rnn_path, segments[curr_segment_idx], curr_segment_idx)
# Get the actual gate position
x_gate_actual = self.env.get_body_pos('gate%d' % curr_segment_idx)
# Calculate the new position from change to new gate position
dx_gate = x_gate_actual - (x_gate_expected + x_path_offset)
x_new = x_next + dx_gate
# Calculate the new velocity
v_new = (x_new - x_curr) / dt_warped
# Calculate the new acceleration
a_new = (v_new - v_curr) / dt_warped
# Calculate the acceleration vector norm
a_new_norm = np.linalg.norm(a_new)
# Limit the norm vector
a_new_norm_clipped = np.clip(a_new_norm, -a_max, a_max)
# Determine the ratio of the clipped norm, protect against divide by zero
if a_new_norm != 0.0:
ratio_unclipped = a_new_norm_clipped / a_new_norm
else:
ratio_unclipped = 0.0
# Scale the acceleration vector by this ratio
a_new = a_new * ratio_unclipped
# Calculate the new change in velocity
dv_new = a_new * dt_warped
v_new = v_curr + dv_new
# Calculate the new change in position
dx_new = v_new * dt_warped
x_new = x_curr + dx_new
# Modify final path data with current tooltip and gate positions
pathutils.set_path_time(final_path, path_idx, t)
pathutils.set_path_tooltip_pos(final_path, path_idx, x_curr)
for gate_idx in range(constants.G_NUM_GATES):
gate_name = 'gate%d' % gate_idx
x_gate = self.env.get_body_pos(gate_name)
pathutils.set_path_gate_pos(final_path, path_idx, gate_idx,
x_gate)
# Store this velocity for the next time step
v_curr = v_new
# Recalculate the current offset
x_path_offset += x_new - x_next
# Perform inverse kinematics to get joint angles
pa10_joint_angles = self.kinematics.calc_inverse_kinematics(
x_curr, x_new)
# TODO: TEMP - MOVE ONLY POINTER, NO PA10
self.env.set_group_pos('pointer', x_new)
if constants.G_TABLE_IS_OSCILLATING:
# Move the table with y-axis oscillation
x_table_next = shaker_table(t, x_table)
else:
x_table_next = x_table
self.env.set_body_pos('table', x_table_next)
# Step through the world by 1 time frame and actuate pa10 joints
self.env.performAction(pa10_joint_angles, fast=fast_step)
# Update current time after this step
t += dt_warped
path_idx += 1
# Determine the difference in virtual vs actual time
t_warped = dt - (time.time() - t_start)
# Attempt to enforce real-time constraints
if t_warped >= 0.0:
# The calculation took less time than the virtual time. Sleep
# the rest off
time.sleep(t_warped)
t_overshoot = 0.0
else:
# The calculation took more time than the virtual time. We need
# to catch up with the virtual time on the next time step
t_overshoot = -t_warped
return
def train_path_planning_network():
"""Train Path Planning Network
Trains an Evolino LSTM neural network for long-term path planning for
use in the surgical simulator.
Returns:
A copy of the fully-trained path planning neural network.
"""
# Build up the list of files to use as training set
training_dir = constants.G_TRAINING_DATA_DIR
# Find all data files in the training data directory
training_files = pathutils.list_data_files(training_dir)
# Get the training data and place it into a dataset
training_dataset = None
# Store all training set ratings
ratings = np.array([])
for training_file in training_files:
training_data = datastore.retrieve(training_file)
# Normalize the time input of the data
training_data = pathutils.normalize_time(training_data, t_col=constants.G_TIME_IDX)
# Add this data sample to the training dataset
training_dataset = datastore.list_to_dataset(
training_data[:,constants.G_RNN_INPUT_IDX:constants.G_RNN_INPUT_IDX+constants.G_RNN_NUM_INPUTS],
training_data[:,constants.G_RNN_OUTPUT_IDX:constants.G_RNN_OUTPUT_IDX+constants.G_RNN_NUM_OUTPUTS],
dataset=training_dataset
)
# Store the rating of the data
this_rating = training_data[1:,constants.G_RATING_IDX]
ratings = np.hstack((ratings, this_rating))
# Get the starting point information for testing
output_start_idx = constants.G_RNN_OUTPUT_IDX
output_end_idx = output_start_idx + constants.G_RNN_NUM_OUTPUTS
output_initial_condition = training_data[0,output_start_idx:output_end_idx]
# Generate the time sequence input data for testing
time_steps = constants.G_RNN_GENERATED_TIME_STEPS
t_input = np.linspace(start=0.0, stop=1.0, num=time_steps)
t_input = np.reshape(t_input, (len(t_input), 1))
gate_start_idx = constants.G_GATE_IDX
gate_end_idx = gate_start_idx + constants.G_NUM_GATE_INPUTS
# Pull the gate data from the last training dataset
gate_data = training_data[0:1,gate_start_idx:gate_end_idx]
gate_data = np.tile(gate_data, (time_steps, 1))
# Build up a full ratings matrix
nd_ratings = None
for rating in ratings:
this_rating = rating * np.ones((1, constants.G_RNN_NUM_OUTPUTS))
if nd_ratings is None:
nd_ratings = this_rating
else:
nd_ratings = np.vstack((nd_ratings, this_rating))
# Create network and trainer
print('>>> Building Network...')
net = PathPlanningNetwork()
print('>>> Initializing Trainer...')
trainer = PathPlanningTrainer(
evolino_network=net,
dataset=training_dataset,
nBurstMutationEpochs=10,
importance=nd_ratings
)
# Begin the training iterations
fitness_list = []
max_fitness = None
max_fitness_epoch = None
# Draw the generated path plot
fig = plt.figure(1, facecolor='white')
testing_axis = fig.add_subplot(111, projection='3d')
fig.show()
idx = 0
current_convergence_streak = 0
while True:
print('>>> Training Network (Iteration: %3d)...' % (idx+1))
trainer.train()
# Determine fitness of this network
current_fitness = trainer.evaluation.max_fitness
fitness_list.append(current_fitness)
print('>>> FITNESS: %f' % current_fitness)
# Determine if this is the minimal error network
if max_fitness is None or max_fitness < current_fitness:
# This is the minimum, record it
max_fitness = current_fitness
max_fitness_epoch = idx
# Draw the generated path after training
print('>>> Testing Network...')
generated_output = net.extrapolate(t_input, [output_initial_condition], len(t_input)-1)
generated_output = np.vstack((output_initial_condition, generated_output))
generated_input = np.hstack((t_input, gate_data))
# Smash together the input and output
generated_data = np.hstack((generated_input, generated_output))
print('>>> Drawing Generated Path...')
pathutils.display_path(testing_axis, generated_data, title='Generated Testing Path')
plt.draw()
if current_fitness > constants.G_RNN_CONVERGENCE_THRESHOLD:
# We've encountered a fitness higher than threshold
current_convergence_streak += 1
else:
# Streak ended. Reset the streak counter
current_convergence_streak = 0
if current_convergence_streak == constants.G_RNN_REQUIRED_CONVERGENCE_STREAK:
print('>>> Convergence Achieved: %d Iterations' % idx)
break
elif idx == constants.G_RNN_MAX_ITERATIONS - 1:
print('>>> Reached maximum iterations (%d)' % constants.G_RNN_MAX_ITERATIONS)
break
idx += 1
# Draw the iteration fitness plot
plt.figure(facecolor='white')
plt.cla()
plt.title('Fitness of RNN over %d Iterations' % (idx-1))
plt.xlabel('Training Iterations')
plt.ylabel('Fitness')
plt.grid(True)
plt.plot(range(len(fitness_list)), fitness_list, 'r-')
plt.annotate('local max', xy=(max_fitness_epoch, fitness_list[max_fitness_epoch]),
xytext=(max_fitness_epoch, fitness_list[max_fitness_epoch]+0.01),
arrowprops=dict(facecolor='black', shrink=0.05))
plt.show()
# Return a full copy of the trained neural network
return copy.deepcopy(net)
def start(self, fps, fast_step=False):
"""Start
Begin the continuous event loop for the simulation. This event loop
can be exited using the ctrl+c keyboard interrupt. Real-time
constraints are enforced. [Hz]
Arguments:
fps: The value of frames per second of the simulation.
fast_step: If True, the ODE fast step algorithm will be used.
This is faster and requires less memory but is less accurate.
(Default: False)
"""
paused = False
stopped = False
# Define the total time for the tooltip traversal
t_total = 20.0
# Define the simulation frame rate
t = 0.0 # [s]
dt = 1.0 / float(fps) # [s]
# Keep track of time overshoot in the case that simulation time must be
# increased in order to maintain real-time constraints
t_overshoot = 0.0
# Get the initial path position (center of gate7)
pos_start = self.env.get_body_pos('gate7') # [m]
# Get the first position of the PA10 at rest
pos_init = self.env.get_body_pos('tooltip') # [m]
# Calculate the new required joint angles of the PA10
#pa10_joint_angles = self.kinematics.calc_inverse_kinematics(pos_init, pos_start)
# TODO: Move the PA10 end-effector to the starting position along the path
# TODO: TEMP - Move the temporary end-effector pointer to the starting position
self.env.set_group_pos('pointer', pos_start)
# Generate long-term path from initial position
t_input = np.linspace(start=0.0, stop=1.0, num=t_total/dt)
t_input = np.reshape(t_input, (len(t_input), 1))
rnn_path = self.rnn.extrapolate(t_input, [pos_start], len(t_input)-1)
# Add the initial condition point back onto the data
rnn_path = np.vstack((pos_start, rnn_path))
# Retrieve one set of standard gate position/orientation data
file_path = pathutils.list_data_files(constants.G_TRAINING_DATA_DIR)[0]
gate_data = datastore.retrieve(file_path)
gate_start_idx = constants.G_GATE_IDX
gate_end_idx = gate_start_idx + constants.G_NUM_GATE_INPUTS
# Reshape the gate positions data
gate_data = gate_data[0:1,gate_start_idx:gate_end_idx]
gate_data = np.tile(gate_data, (len(rnn_path), 1))
# Complete the rnn path data
rnn_path = np.hstack((t_input, gate_data, rnn_path))
# Save generated path for later examination
datastore.store(rnn_path, constants.G_RNN_STATIC_PATH_OUT)
# Define a variable to hold the final path (with real-time correction)
final_path = rnn_path[:-1].copy()
path_saved = False
# Detect all path segments between gates in the generated path
segments = pathutils._detect_segments(rnn_path)
path_idx = 0
x_path_offset = np.array([0.0, 0.0, 0.0]) # [m]
v_curr = np.array([0.0, 0.0, 0.0]) # [m/s]
a_max = constants.G_MAX_ACCEL # [m/s^2]
# Get the static table position
x_table = self.env.get_body_pos('table')
while not stopped:
t_start = time.time()
# If the last calculation took too long, catch up
dt_warped = dt + t_overshoot
self.env.set_dt(dt_warped)
# Determine if the viewer is stopped. Then we can quit
if self.viewer.is_dead:
break
# Pause the simulation if we are at the end
if path_idx == len(rnn_path) - 1 or paused:
self.env.step(paused=True, fast=fast_step)
# If we have really hit the end of the simulation, save/plot the path
if not paused and not path_saved:
# Save the final data to a file
datastore.store(final_path, constants.G_RNN_DYNAMIC_PATH_OUT)
path_saved = True
continue
# Not a very elegant solution to pausing at the start, but it works
if t <= 1000.0:
self.env.step(paused=True, fast=fast_step)
t += dt_warped
continue
# Determine the current path segment
curr_segment_idx = 0
for segment_idx, segment_end in enumerate(segments):
if path_idx <= segment_end:
curr_segment_idx = segment_idx
break
x_curr = pathutils.get_path_tooltip_pos(rnn_path, path_idx) + x_path_offset
x_next = pathutils.get_path_tooltip_pos(rnn_path, path_idx+1) + x_path_offset
# Get the expected gate position
x_gate_expected = pathutils.get_path_gate_pos(
rnn_path,
segments[curr_segment_idx],
curr_segment_idx
)
# Get the actual gate position
x_gate_actual = self.env.get_body_pos('gate%d'%curr_segment_idx)
# Calculate the new position from change to new gate position
dx_gate = x_gate_actual - (x_gate_expected + x_path_offset)
x_new = x_next + dx_gate
# Calculate the new velocity
v_new = (x_new - x_curr) / dt_warped
# Calculate the new acceleration
a_new = (v_new - v_curr) / dt_warped
# Calculate the acceleration vector norm
a_new_norm = np.linalg.norm(a_new)
# Limit the norm vector
a_new_norm_clipped = np.clip(a_new_norm, -a_max, a_max)
# Determine the ratio of the clipped norm, protect against divide by zero
if a_new_norm != 0.0:
ratio_unclipped = a_new_norm_clipped / a_new_norm
else:
ratio_unclipped = 0.0
# Scale the acceleration vector by this ratio
a_new = a_new * ratio_unclipped
# Calculate the new change in velocity
dv_new = a_new * dt_warped
v_new = v_curr + dv_new
# Calculate the new change in position
dx_new = v_new * dt_warped
x_new = x_curr + dx_new
# Modify final path data with current tooltip and gate positions
pathutils.set_path_time(final_path, path_idx, t)
pathutils.set_path_tooltip_pos(final_path, path_idx, x_curr)
for gate_idx in range(constants.G_NUM_GATES):
gate_name = 'gate%d' % gate_idx
x_gate = self.env.get_body_pos(gate_name)
pathutils.set_path_gate_pos(final_path, path_idx, gate_idx, x_gate)
# Store this velocity for the next time step
v_curr = v_new
# Recalculate the current offset
x_path_offset += x_new - x_next
# Perform inverse kinematics to get joint angles
pa10_joint_angles = self.kinematics.calc_inverse_kinematics(x_curr, x_new)
# TODO: TEMP - MOVE ONLY POINTER, NO PA10
self.env.set_group_pos('pointer', x_new)
if constants.G_TABLE_IS_OSCILLATING:
# Move the table with y-axis oscillation
x_table_next = shaker_table(t, x_table)
else:
x_table_next = x_table
self.env.set_body_pos('table', x_table_next)
# Step through the world by 1 time frame and actuate pa10 joints
self.env.performAction(pa10_joint_angles, fast=fast_step)
# Update current time after this step
t += dt_warped
path_idx += 1
# Determine the difference in virtual vs actual time
t_warped = dt - (time.time() - t_start)
# Attempt to enforce real-time constraints
if t_warped >= 0.0:
# The calculation took less time than the virtual time. Sleep
# the rest off
time.sleep(t_warped)
t_overshoot = 0.0
else:
# The calculation took more time than the virtual time. We need
# to catch up with the virtual time on the next time step
t_overshoot = -t_warped
return
