def build_training_graph(num_agents):
s = tf.placeholder(tf.float32, [num_agents, 4])
g = tf.placeholder(tf.float32, [num_agents, 2])
x = tf.expand_dims(s, 1) - tf.expand_dims(s, 0)
h, mask, indices = core.network_cbf(x=x, r=config.DIST_MIN_THRES, indices=None)
a = core.network_action(s=s, g=g, obs_radius=config.OBS_RADIUS, indices=indices)
(loss_dang, loss_safe, acc_dang, acc_safe) = core.loss_barrier(
h=h, s=s, r=config.DIST_MIN_THRES, ttc=config.TIME_TO_COLLISION, indices=indices)
(loss_dang_deriv, loss_safe_deriv, acc_dang_deriv, acc_safe_deriv
) = core.loss_derivatives(s=s, a=a, h=h, x=x, r=config.DIST_MIN_THRES,
indices=indices, ttc=config.TIME_TO_COLLISION, alpha=config.ALPHA_CBF)
loss_action = core.loss_actions(
s=s, g=g, a=a, r=config.DIST_MIN_THRES, ttc=config.TIME_TO_COLLISION)
loss_list = [2 * loss_dang, loss_safe, 2 * loss_dang_deriv, loss_safe_deriv, 0.01 * loss_action]
acc_list = [acc_dang, acc_safe, acc_dang_deriv, acc_safe_deriv]
weight_loss = [
config.WEIGHT_DECAY * tf.nn.l2_loss(v) for v in tf.trainable_variables()]
loss = 10 * tf.math.add_n(loss_list + weight_loss)
return s, g, a, loss_list, loss, acc_list
def build_training_graph(num_agents):
# s is the state vectors of the agents
s = tf.placeholder(tf.float32, [num_agents, 8])
# s_ref is the goal states
s_ref = tf.placeholder(tf.float32, [num_agents, 8])
# x is difference between the state of each agent and other agents
x = tf.expand_dims(s, 1) - tf.expand_dims(s, 0)
# h is the CBF value of shape [num_agents, TOP_K, 1], where TOP_K represents
# the K nearest agents
h, mask, indices = core.network_cbf(x=x,
r=config.DIST_MIN_THRES,
indices=None)
# u is the control action of each agent, with shape [num_agents, 3]
u = core.network_action(s=s,
s_ref=s_ref,
obs_radius=config.OBS_RADIUS,
indices=indices)
# compute the value of loss functions and the accuracies
# loss_dang is for h(s) < 0, s in dangerous set
# loss safe is for h(s) >=0, s in safe set
# acc_dang is the accuracy that h(s) < 0, s in dangerous set is satisfied
# acc_safe is the accuracy that h(s) >=0, s in safe set is satisfied
loss_dang, loss_safe, acc_dang, acc_safe = core.loss_barrier(
h=h, s=s, indices=indices)
# loss_dang_deriv is for doth(s) + alpha h(s) >=0 for s in dangerous set
# loss_safe_deriv is for doth(s) + alpha h(s) >=0 for s in safe set
# loss_medium_deriv is for doth(s) + alpha h(s) >=0 for s not in the dangerous
# or the safe set
(loss_dang_deriv, loss_safe_deriv, loss_medium_deriv, acc_dang_deriv,
acc_safe_deriv, acc_medium_deriv) = core.loss_derivatives(s=s,
u=u,
h=h,
x=x,
indices=indices)
# the distance between the u and the nominal u
loss_action = core.loss_actions(s=s, u=u, s_ref=s_ref, indices=indices)
# the weight of each loss item requires careful tuning
loss_list = [
loss_dang, loss_safe, 3 * loss_dang_deriv, loss_safe_deriv,
2 * loss_medium_deriv, 0.5 * loss_action
]
acc_list = [
acc_dang, acc_safe, acc_dang_deriv, acc_safe_deriv, acc_medium_deriv
]
weight_loss = [
config.WEIGHT_DECAY * tf.nn.l2_loss(v)
for v in tf.trainable_variables()
]
loss = 10 * tf.math.add_n(loss_list + weight_loss)
return s, s_ref, u, loss_list, loss, acc_list
def opt_body(u_res, loop_count, is_safe):
dsdt = core.quadrotor_dynamics_tf(s, u + u_res)
s_next = s + dsdt * config.TIME_STEP_EVAL
x_next = tf.expand_dims(s_next, 1) - tf.expand_dims(s_next, 0)
h_next, mask_next, _ = core.network_cbf(x=x_next,
r=config.DIST_MIN_THRES,
indices=indices)
deriv = h_next - h + config.TIME_STEP_EVAL * config.ALPHA_CBF * h
deriv = deriv * mask * mask_next
error = tf.reduce_sum(tf.math.maximum(-deriv, 0), axis=1)
error_gradient = tf.gradients(error, u_res)[0]
u_res = u_res - config.REFINE_LEARNING_RATE * error_gradient
loop_count = loop_count + 1
return u_res, loop_count, is_safe
def opt_body(a_res, loop_count):
dsdt = core.dynamics(s, a + a_res)
s_next = s + dsdt * config.TIME_STEP
x_next = tf.expand_dims(s_next, 1) - tf.expand_dims(s_next, 0)
h_next, mask_next, _ = core.network_cbf(x=x_next,
r=config.DIST_MIN_THRES,
indices=indices)
deriv = h_next - h + config.TIME_STEP * config.ALPHA_CBF * h
deriv = deriv * mask * mask_next
error = tf.reduce_sum(tf.math.maximum(-deriv, 0), axis=1)
error_gradient = tf.gradients(error, a_res)[0]
a_res = a_res - config.REFINE_LEARNING_RATE * error_gradient
loop_count = loop_count + 1
return a_res, loop_count
def opt_body(a_res, loop_count):
# a loop of updating a_res
# compute s_next under a + a_res
dsdt = core.dynamics(s, a + a_res)
s_next = s + dsdt * config.TIME_STEP
x_next = tf.expand_dims(s_next, 1) - tf.expand_dims(s_next, 0)
h_next, mask_next, _ = core.network_cbf(
x=x_next, r=config.DIST_MIN_THRES, indices=indices)
# deriv should be >= 0. if not, we update a_res by gradient descent
deriv = h_next - h + config.TIME_STEP * config.ALPHA_CBF * h
deriv = deriv * mask * mask_next
error = tf.reduce_sum(tf.math.maximum(-deriv, 0), axis=1)
# compute the gradient to update a_res
error_gradient = tf.gradients(error, a_res)[0]
a_res = a_res - config.REFINE_LEARNING_RATE * error_gradient
loop_count = loop_count + 1
return a_res, loop_count
def build_training_graph(num_agents):
s = tf.placeholder(tf.float32, [num_agents, 8])
s_ref = tf.placeholder(tf.float32, [num_agents, 8])
x = tf.expand_dims(s, 1) - tf.expand_dims(s, 0)
h, mask, indices = core.network_cbf(x=x,
r=config.DIST_MIN_THRES,
indices=None)
u = core.network_action(s=s,
s_ref=s_ref,
obs_radius=config.OBS_RADIUS,
indices=indices)
loss_dang, loss_safe, acc_dang, acc_safe = core.loss_barrier(
h=h, s=s, indices=indices)
(loss_dang_deriv, loss_safe_deriv, loss_medium_deriv, acc_dang_deriv,
acc_safe_deriv, acc_medium_deriv) = core.loss_derivatives(s=s,
u=u,
h=h,
x=x,
indices=indices)
loss_action = core.loss_actions(s=s, u=u, s_ref=s_ref, indices=indices)
# the weight of each loss item requires careful tuning
loss_list = [
loss_dang, loss_safe, 3 * loss_dang_deriv, loss_safe_deriv,
2 * loss_medium_deriv, 0.5 * loss_action
]
acc_list = [
acc_dang, acc_safe, acc_dang_deriv, acc_safe_deriv, acc_medium_deriv
]
weight_loss = [
config.WEIGHT_DECAY * tf.nn.l2_loss(v)
for v in tf.trainable_variables()
]
loss = 10 * tf.math.add_n(loss_list + weight_loss)
return s, s_ref, u, loss_list, loss, acc_list
def build_evaluation_graph(num_agents):
s = tf.placeholder(tf.float32, [num_agents, 8])
s_ref = tf.placeholder(tf.float32, [num_agents, 8])
x = tf.expand_dims(s, 1) - tf.expand_dims(s, 0)
h, mask, indices = core.network_cbf(x=x,
r=config.DIST_MIN_THRES,
indices=None)
u = core.network_action(s=s,
s_ref=s_ref,
obs_radius=config.OBS_RADIUS,
indices=indices)
safe_mask = core.compute_safe_mask(s, r=config.DIST_SAFE, indices=indices)
is_safe = tf.equal(tf.reduce_mean(tf.cast(safe_mask, tf.float32)), 1)
u_res = tf.Variable(tf.zeros_like(u), name='u_res')
loop_count = tf.Variable(0, name='loop_count')
def opt_body(u_res, loop_count, is_safe):
dsdt = core.quadrotor_dynamics_tf(s, u + u_res)
s_next = s + dsdt * config.TIME_STEP_EVAL
x_next = tf.expand_dims(s_next, 1) - tf.expand_dims(s_next, 0)
h_next, mask_next, _ = core.network_cbf(x=x_next,
r=config.DIST_MIN_THRES,
indices=indices)
deriv = h_next - h + config.TIME_STEP_EVAL * config.ALPHA_CBF * h
deriv = deriv * mask * mask_next
error = tf.reduce_sum(tf.math.maximum(-deriv, 0), axis=1)
error_gradient = tf.gradients(error, u_res)[0]
u_res = u_res - config.REFINE_LEARNING_RATE * error_gradient
loop_count = loop_count + 1
return u_res, loop_count, is_safe
def opt_cond(u_res, loop_count, is_safe):
cond = tf.logical_and(tf.less(loop_count, config.REFINE_LOOPS),
tf.logical_not(is_safe))
return cond
with tf.control_dependencies(
[u_res.assign(tf.zeros_like(u)),
loop_count.assign(0)]):
u_res, _, _ = tf.while_loop(opt_cond, opt_body,
[u_res, loop_count, is_safe])
u_opt = u + u_res
loss_dang, loss_safe, acc_dang, acc_safe = core.loss_barrier(
h=h, s=s, indices=indices)
(loss_dang_deriv, loss_safe_deriv, loss_medium_deriv, acc_dang_deriv,
acc_safe_deriv, acc_medium_deriv) = core.loss_derivatives(s=s,
u=u_opt,
h=h,
x=x,
indices=indices)
loss_action = core.loss_actions(s=s, u=u_opt, s_ref=s_ref, indices=indices)
loss_list = [
loss_dang, loss_safe, loss_dang_deriv, loss_safe_deriv,
loss_medium_deriv, loss_action
]
acc_list = [
acc_dang, acc_safe, acc_dang_deriv, acc_safe_deriv, acc_medium_deriv
]
return s, s_ref, u_opt, loss_list, acc_list
def build_evaluation_graph(num_agents):
# s is the state vectors of the agents
s = tf.placeholder(tf.float32, [num_agents, 4])
# g is the goal states
g = tf.placeholder(tf.float32, [num_agents, 2])
# x is difference between the state of each agent and other agents
x = tf.expand_dims(s, 1) - tf.expand_dims(s, 0)
# h is the CBF value of shape [num_agents, TOP_K, 1], where TOP_K represents
# the K nearest agents
h, mask, indices = core.network_cbf(x=x, r=config.DIST_MIN_THRES)
# a is the control action of each agent, with shape [num_agents, 3]
a = core.network_action(s=s, g=g, obs_radius=config.OBS_RADIUS, indices=indices)
# a_res is delta a. when a does not satisfy the CBF conditions, we want to compute
# a a_res such that a + a_res satisfies the CBF conditions
a_res = tf.Variable(tf.zeros_like(a), name='a_res')
loop_count = tf.Variable(0, name='loop_count')
def opt_body(a_res, loop_count):
# a loop of updating a_res
# compute s_next under a + a_res
dsdt = core.dynamics(s, a + a_res)
s_next = s + dsdt * config.TIME_STEP
x_next = tf.expand_dims(s_next, 1) - tf.expand_dims(s_next, 0)
h_next, mask_next, _ = core.network_cbf(
x=x_next, r=config.DIST_MIN_THRES, indices=indices)
# deriv should be >= 0. if not, we update a_res by gradient descent
deriv = h_next - h + config.TIME_STEP * config.ALPHA_CBF * h
deriv = deriv * mask * mask_next
error = tf.reduce_sum(tf.math.maximum(-deriv, 0), axis=1)
# compute the gradient to update a_res
error_gradient = tf.gradients(error, a_res)[0]
a_res = a_res - config.REFINE_LEARNING_RATE * error_gradient
loop_count = loop_count + 1
return a_res, loop_count
def opt_cond(a_res, loop_count):
# update u_res for REFINE_LOOPS
cond = tf.less(loop_count, config.REFINE_LOOPS)
return cond
with tf.control_dependencies([
a_res.assign(tf.zeros_like(a)), loop_count.assign(0)]):
a_res, _ = tf.while_loop(opt_cond, opt_body, [a_res, loop_count])
a_opt = a + a_res
dsdt = core.dynamics(s, a_opt)
s_next = s + dsdt * config.TIME_STEP
x_next = tf.expand_dims(s_next, 1) - tf.expand_dims(s_next, 0)
h_next, mask_next, _ = core.network_cbf(x=x_next, r=config.DIST_MIN_THRES, indices=indices)
# compute the value of loss functions and the accuracies
# loss_dang is for h(s) < 0, s in dangerous set
# loss safe is for h(s) >=0, s in safe set
# acc_dang is the accuracy that h(s) < 0, s in dangerous set is satisfied
# acc_safe is the accuracy that h(s) >=0, s in safe set is satisfied
(loss_dang, loss_safe, acc_dang, acc_safe) = core.loss_barrier(
h=h_next, s=s_next, r=config.DIST_MIN_THRES,
ttc=config.TIME_TO_COLLISION, eps=[0, 0])
# loss_dang_deriv is for doth(s) + alpha h(s) >=0 for s in dangerous set
# loss_safe_deriv is for doth(s) + alpha h(s) >=0 for s in safe set
# loss_medium_deriv is for doth(s) + alpha h(s) >=0 for s not in the dangerous
# or the safe set
(loss_dang_deriv, loss_safe_deriv, acc_dang_deriv, acc_safe_deriv
) = core.loss_derivatives(s=s_next, a=a_opt, h=h_next, x=x_next,
r=config.DIST_MIN_THRES, ttc=config.TIME_TO_COLLISION, alpha=config.ALPHA_CBF, indices=indices)
# the distance between the u_opt and the nominal u
loss_action = core.loss_actions(s, g, a, r=config.DIST_MIN_THRES, ttc=config.TIME_TO_COLLISION)
loss_list = [loss_dang, loss_safe, loss_dang_deriv, loss_safe_deriv, loss_action]
acc_list = [acc_dang, acc_safe, acc_dang_deriv, acc_safe_deriv]
return s, g, a_opt, loss_list, acc_list
def build_evaluation_graph(num_agents):
s = tf.placeholder(tf.float32, [num_agents, 4])
g = tf.placeholder(tf.float32, [num_agents, 2])
x = tf.expand_dims(s, 1) - tf.expand_dims(s, 0)
h, mask, indices = core.network_cbf(x=x, r=config.DIST_MIN_THRES)
a = core.network_action(s=s,
g=g,
obs_radius=config.OBS_RADIUS,
indices=indices)
a_res = tf.Variable(tf.zeros_like(a), name='a_res')
loop_count = tf.Variable(0, name='loop_count')
def opt_body(a_res, loop_count):
dsdt = core.dynamics(s, a + a_res)
s_next = s + dsdt * config.TIME_STEP
x_next = tf.expand_dims(s_next, 1) - tf.expand_dims(s_next, 0)
h_next, mask_next, _ = core.network_cbf(x=x_next,
r=config.DIST_MIN_THRES,
indices=indices)
deriv = h_next - h + config.TIME_STEP * config.ALPHA_CBF * h
deriv = deriv * mask * mask_next
error = tf.reduce_sum(tf.math.maximum(-deriv, 0), axis=1)
error_gradient = tf.gradients(error, a_res)[0]
a_res = a_res - config.REFINE_LEARNING_RATE * error_gradient
loop_count = loop_count + 1
return a_res, loop_count
def opt_cond(a_res, loop_count):
cond = tf.less(loop_count, config.REFINE_LOOPS)
return cond
with tf.control_dependencies(
[a_res.assign(tf.zeros_like(a)),
loop_count.assign(0)]):
a_res, _ = tf.while_loop(opt_cond, opt_body, [a_res, loop_count])
a_opt = a + a_res
dsdt = core.dynamics(s, a_opt)
s_next = s + dsdt * config.TIME_STEP
x_next = tf.expand_dims(s_next, 1) - tf.expand_dims(s_next, 0)
h_next, mask_next, _ = core.network_cbf(x=x_next,
r=config.DIST_MIN_THRES,
indices=indices)
(loss_dang, loss_safe, acc_dang,
acc_safe) = core.loss_barrier(h=h_next,
s=s_next,
r=config.DIST_MIN_THRES,
ttc=config.TIME_TO_COLLISION,
eps=[0, 0])
(loss_dang_deriv, loss_safe_deriv, acc_dang_deriv,
acc_safe_deriv) = core.loss_derivatives(s=s_next,
a=a_opt,
h=h_next,
x=x_next,
r=config.DIST_MIN_THRES,
ttc=config.TIME_TO_COLLISION,
alpha=config.ALPHA_CBF,
indices=indices)
loss_action = core.loss_actions(s,
g,
a,
r=config.DIST_MIN_THRES,
ttc=config.TIME_TO_COLLISION)
loss_list = [
loss_dang, loss_safe, loss_dang_deriv, loss_safe_deriv, loss_action
]
acc_list = [acc_dang, acc_safe, acc_dang_deriv, acc_safe_deriv]
return s, g, a_opt, loss_list, acc_list
def build_training_graph(num_agents):
# s is the state vectors of the agents
s = tf.placeholder(tf.float32, [num_agents, 4])
# g is the goal states
g = tf.placeholder(tf.float32, [num_agents, 2])
# x is difference between the state of each agent and other agents
x = tf.expand_dims(s, 1) - tf.expand_dims(s, 0)
# h is the CBF value of shape [num_agents, TOP_K, 1], where TOP_K represents
# the K nearest agents
h, mask, indices = core.network_cbf(x=x,
r=config.DIST_MIN_THRES,
indices=None)
# a is the control action of each agent, with shape [num_agents, 2]
a = core.network_action(s=s,
g=g,
obs_radius=config.OBS_RADIUS,
indices=indices)
# compute the value of loss functions and the accuracies
# loss_dang is for h(s) < 0, s in dangerous set
# loss safe is for h(s) >=0, s in safe set
# acc_dang is the accuracy that h(s) < 0, s in dangerous set is satisfied
# acc_safe is the accuracy that h(s) >=0, s in safe set is satisfied
(loss_dang, loss_safe, acc_dang,
acc_safe) = core.loss_barrier(h=h,
s=s,
r=config.DIST_MIN_THRES,
ttc=config.TIME_TO_COLLISION,
indices=indices)
# loss_dang_deriv is for doth(s) + alpha h(s) >=0 for s in dangerous set
# loss_safe_deriv is for doth(s) + alpha h(s) >=0 for s in safe set
# loss_medium_deriv is for doth(s) + alpha h(s) >=0 for s not in the dangerous
# or the safe set
(loss_dang_deriv, loss_safe_deriv, acc_dang_deriv,
acc_safe_deriv) = core.loss_derivatives(s=s,
a=a,
h=h,
x=x,
r=config.DIST_MIN_THRES,
indices=indices,
ttc=config.TIME_TO_COLLISION,
alpha=config.ALPHA_CBF)
# the distance between the a and the nominal a
loss_action = core.loss_actions(s=s,
g=g,
a=a,
r=config.DIST_MIN_THRES,
ttc=config.TIME_TO_COLLISION)
loss_list = [
2 * loss_dang, loss_safe, 2 * loss_dang_deriv, loss_safe_deriv,
0.01 * loss_action
]
acc_list = [acc_dang, acc_safe, acc_dang_deriv, acc_safe_deriv]
weight_loss = [
config.WEIGHT_DECAY * tf.nn.l2_loss(v)
for v in tf.trainable_variables()
]
loss = 10 * tf.math.add_n(loss_list + weight_loss)
return s, g, a, loss_list, loss, acc_list