Interacting with an Experiment from the Virtual Coach¶
In the previous section we successfully launched a simulation of the Template Husky experiment. In this tutorial, we will interact with the simulation through the Virtual Coach.
Getting and Setting Simulation States¶
When an experiment is launched, it is initially in the paused state. This means the experiment has to be explicitly started. In general, we can query the simulation state by calling:
The state of the simulation is returned as a string. In this case calling
get_state() will return paused. The possible states we can set a simulation to are paused, started and stopped. We can alternatively start and pause the simulation by calling:
Note that the timeout for a simulation is 15 minutes.
From the Virtual Coach, we can view transfer functions, modify them, delete them, or add new ones. The transfer functions unique identifiers are their function names. To know the names of the transfer functions defined in an experiments, we can call:
which would just print out the names of the transfer functions. In the case of the Template Husky experiment, this call will print out: turn_around, csv_spike_monitor, all_neurons_spike_monitor, grab_image, and csv_joint_state_monitor. To actually see the code of a transfer function, we can use the get_transfer_function call, which takes the transfer function name as a parameter and returns the code in a string. We will get the code of the turn-around transfer function and store it in a variable:
turn_around_tf = sim.get_transfer_function('turn_around')
If you’re using a jupyter notebook, you can use the
%load command to load the transfer function code in a cell. This will allow you to see the code syntax highlighted and properly indented and will make it easier to modify the functions:
This should load this function in a new cell in the jupyter notebook:
import hbp_nrp_cle.tf_framework as nrp from hbp_nrp_cle.robotsim.RobotInterface import Topic import geometry_msgs.msg @nrp.MapSpikeSink("output_neuron", nrp.brain.neurons, nrp.leaky_integrator_alpha) @nrp.Neuron2Robot(Topic('/husky/husky/cmd_vel', geometry_msgs.msg.Twist)) # Example TF: get output neuron voltage and output constant on robot actuator. You could do something with the voltage here and command the robot accordingly. def turn_around(t, output_neuron): voltage=output_neuron.voltage return geometry_msgs.msg.Twist(linear=geometry_msgs.msg.Vector3(0,0,0), angular=geometry_msgs.msg.Vector3(0,0,5))
You can modify the code however you want and then save it as a string. Here we’ll just increase the angular velocity by modifying the last line in the above code and replacing the number 5 with 10. We will save the modified code as a string in the variable turn_around_tf:
turn_around_tf = """ import hbp_nrp_cle.tf_framework as nrp from hbp_nrp_cle.robotsim.RobotInterface import Topic import geometry_msgs.msg @nrp.MapSpikeSink("output_neuron", nrp.brain.neurons, nrp.leaky_integrator_alpha) @nrp.Neuron2Robot(Topic('/husky/husky/cmd_vel', geometry_msgs.msg.Twist)) # Example TF: get output neuron voltage and output constant on robot actuator. You could do something with the voltage here and command the robot accordingly. def turn_around(t, output_neuron): voltage=output_neuron.voltage return geometry_msgs.msg.Twist(linear=geometry_msgs.msg.Vector3(0,0,0), angular=geometry_msgs.msg.Vector3(0,0,10)) """
This modified transfer function will only make the robot spin faster in this experiment. If you open your frontend web cockpit and join the running experiment, you will see the robot spinning faster once we actually apply the transfer function. To apply the transfer function we use the call edit_transfer_function which takes as parameters the name of the transfer function to be modified and the modified code.
The Virtual Coach will maintain the simulation state after setting the transfer function. This means that if the simulation was running, the Virtual Coach will modify the transfer function and then automatically start the simulation again.
As a user you can also delete transfer functions from the Virtual Coach. You just need to provide the name of the transfer function and use it in the following call:
This will delete the turn_around transfer function we just modified. If you have a frontend web cockpit joined on your simulation, you will notice that the robot stopped spinning since the transfer function responsible for that behavior has been deleted. Note that deletion and addition of transfer functions are not reflected in the frontend. If you want more proof that the transfer function has been deleted, you can revisit the print_transfer_functions call and make sure that it doesn’t print out turn_around.
We can also add new transfer functions. For this we only need to provide the transfer function code as a string parameter to the add_transfer_function function. We don’t have to provide a name since the name will just be the function’s name. Remember that transfer functions definition names have to be unique, so duplicate function names will result in errors. Here we’ll create three transfer functions that store Spike, Joint and Robot positions into csv files.
csv_spike_monitor = """@nrp.MapCSVRecorder("recorder", filename="all_spikes.csv", headers=["id", "time"]) @nrp.MapSpikeSink("record_neurons", nrp.brain.record, nrp.spike_recorder) @nrp.Neuron2Robot(Topic('/monitor/spike_recorder', cle_ros_msgs.msg.SpikeEvent)) def csv_spike_monitor(t, recorder, record_neurons): for i in range(0, len(record_neurons.times)): recorder.record_entry( record_neurons.times[i], record_neurons.times[i] )""" sim.add_transfer_function(csv_spike_monitor)
csv_joint_state_monitor = """@nrp.MapRobotSubscriber("joint_state", Topic('/husky/joint_states', sensor_msgs.msg.JointState)) @nrp.MapCSVRecorder("recorder", filename="all_joints_positions.csv", headers=["Name", "time", "Position"]) def csv_joint_state_monitor(t, joint_state, recorder): if not isinstance(joint_state.value, type(None)): for i in range(0, len(joint_state.value.name)): recorder.record_entry(joint_state.value.name[i], t, joint_state.value.position[i])""" sim.add_transfer_function(csv_joint_state_monitor)
csv_robot_position = """@nrp.MapCSVRecorder("recorder", filename="robot_position.csv", headers=["x", "y", "z"]) @nrp.MapRobotSubscriber("position", Topic('/gazebo/model_states', gazebo_msgs.msg.ModelStates)) @nrp.MapVariable("robot_index", global_key="robot_index", initial_value=None) @nrp.Robot2Neuron() def csv_robot_position(t, position, recorder, robot_index): if not isinstance(position.value, type(None)): # determine if previously set robot index has changed if robot_index.value is not None: # if the value is invalid, reset the index below if robot_index.value >= len(position.value.name) or\ position.value.name[robot_index.value] != 'husky': robot_index.value = None # robot index is invalid, find and set it if robot_index.value is None: # 'husky' is the bodyModel declared in the bibi, if not found raise error robot_index.value = position.value.name.index('husky') # record the current robot position recorder.record_entry(position.value.pose[robot_index.value].position.x, position.value.pose[robot_index.value].position.y, position.value.pose[robot_index.value].position.z)""" sim.add_transfer_function(csv_robot_position)
Those transfer functions will log the simulation time to the log console every two seconds.
Getting Experiment Data¶
With the transfer functions that we wrote, we can access all experiment data from the Virtual Coach and plot or analyze the data. You can specify, if you want to inspect ‘csv’ data or ‘profiler’ data, first is created with Transfer Function recording, latter from the NRP cle profiler. To know what kind of data is being saved to files in an experiment, you can print out the names of the files first using this call:
In the case of the Template Husky experiment, this will print out all_spikes.csv and all_joints_positions.csv and robot_position.csv. We can now get the data from any one of these files. Note that these files will be populated only if a simulation has been running. Here we will get the Spike data:
spikes = vc.get_last_run_file(exp_id, 'csv', 'all_spikes.csv') print(spikes) [[u'id', u'time', u'Simulation_reset'], [u'3.0', u'0.10000000000000001', u'RESET'], [u'4.0', u'2.6000000000000001', u''], [u'3.0', u'57.200000000000003', u'']]
In the code snippet above you can notice the additional Simulation_reset column which automatically keeps track of reset events.
We can also write our own custom functions to plot the data we got. The following is a custom function that will plot each spike from the csv file as a blue dot. Note also that the first line in the csv data is a header that need to be accounted for when plotting.
from StringIO import StringIO import pandas as pd import matplotlib.pyplot as plt spikes_df = pd.read_csv(StringIO(spikes), sep=",") spikes_df.plot.scatter('time','id') plt.show()
Through the Virtual Coach, users can interact with the simulation state machines the same way they can with the transfer functions. Currently we have only one experiment that contains a state machine. Let’s stop our current simulation and start it and see how we can interact with the state machines.
sim.stop() exp_id = vc.clone_experiment_to_storage('ScreenSwitchingHuskyExperiment') sim = vc.launch_experiment(exp_id)
After the experiment has been started, we can retrieve the names of the defined state machines.
This call should print out HuskyAwareScreenControlling. To retrieve the code of the state machine, we will have to use its name we just got.
sm = sim.get_state_machine('HuskyAwareScreenControlling')
Since state machines are also python scripts, we can load them in jupyter notebooks with the
%load command like we did with the transfer functions. Additionally, we can also edit and delete them, or add new ones, exactly like we interact with transfer functions. Below are the calls for editing, deleting and adding state machines.
sim.edit_state_machine(state_machine_name, state_machine_code) sim.delete_state_machine(state_machine_name) sim.add_state_machine(state_machine_name, state_machine_code)
The only difference between interacting with state machines and transfer functions is that the state machines’ are not the python function names. Therefore, when adding a new state machine, the user has to explicitly give it a name.
It is also possible to reset certain aspects of the simulation from the Virtual Coach, exactly as it is possible from the web cockpit. There are four reset types possible from the Virtual Coach: Robot Frame, Environment, Brain, and the Full Simulation. You can reset all simulation aspects with the same call:
sim.reset('robot_pose') sim.reset('world') sim.reset('brain') sim.reset('full')
If you want to look at more concrete example experiments run from the Virtual Coach, you can check out the hbp_nrp_virtual_coach/examples directory.