This guide and guide_hardware are lightly out of date with v0.4.0 of autopilot, but still largely reflect the program design and its operation. For a simpler task, see Blink .

Many of these things can be done more elegantly, more simply, etc. now but we are a very small team and can only do so much work between releases! We’d be happy to get documentation requests or even a pull request or two to help us out until we can get to it :)

Writing a Task

Some concepts of task design are also discussed in section 3.1 of the whitepaper.

The Nafc Task

The Nafc class serves as an example for new task designs.

To demonstrate the general structure of Autopilot tasks, let’s build it from scratch.

The Task class

We start by subclassing the Task class and initializing it.

from autopilot.tasks import Task

class Nafc(Task):

    def __init__(self):
        super(Nafc, self).__init__()

This gives our new task some basic attributes and methods, including the init_hardware() method for initializing the HARDWARE dictionary and the handle_trigger() method for handling GPIO triggers.

Four Task Attributes

We then add the four elements of a task description:

  1. A PARAMS dictionary defines what parameters are needed to define the task

  2. A Data (tables.IsDescription) descriptor describes what data will be returned from the task

  3. A PLOT dictionary that maps the data output to graphical elements in the GUI.

  4. A HARDWARE dictionary that describes what hardware will be needed to run the task.


Each parameter needs a human readable tag that will be used for GUI elements, and a type, currently one of:

  • int: integers

  • bool: boolean (checkboxes in GUI)

  • list: list of possible values in {‘Name’:int} pairs

  • sounds: a autopilot.core.gui.Sound_Widget to define sounds.

To maintain order when opened by the GUI we use a odict rather than a normal dictionary.

from collections import odict

PARAMS = odict()
PARAMS['reward']         = {'tag':'Reward Duration (ms)',
PARAMS['req_reward']     = {'tag':'Request Rewards',
PARAMS['punish_stim']    = {'tag':'White Noise Punishment',
PARAMS['punish_dur']     = {'tag':'Punishment Duration (ms)',
PARAMS['correction']     = {'tag':'Correction Trials',
PARAMS['correction_pct'] = {'tag':'% Correction Trials',
PARAMS['bias_mode']      = {'tag':'Bias Correction Mode',
                                      'Thresholded Proportional':2}}
PARAMS['bias_threshold'] = {'tag': 'Bias Correction Threshold (%)',
PARAMS['stim']           = {'tag':'Sounds',


See the Nafc class for descriptions of the task parameters.

These will be taken as key-value pairs when the task is initialized. ie.:

PARAMS['correction'] = {'tag':  'Correction Trials',
                        'type': 'bool'}

will be used to initialize the task like:

Nafc(correction=True) # or False


There are two types of data,

  • TrialData - where a single value for several variables is returned per ‘trial’, and

  • ContinuousData - where values and timestamps are taken continuously, with either a fixed or variable interval

Both are defined by pytables tables.IsDescription objects. Specify each variable that will be returned and its type using a tables.Col object:


See the pytables documentation for a list of Col types

import tables

class TrialData(tables.IsDescription):
    trial_num    = tables.Int32Col()
    target       = tables.StringCol(1)
    response     = tables.StringCol(1)
    correct      = tables.Int32Col()
    correction   = tables.Int32Col()
    RQ_timestamp = tables.StringCol(26)
    DC_timestamp = tables.StringCol(26)
    bailed       = tables.Int32Col()

The column types are names with their type and their bit depth except for the StringCol which takes a string length in characters.

The TrialData object is used by the Subject class when a task is assigned to create the data storage table.


The PLOT dictionary maps the data returned from the task to graphical elements in the Terminal’s Plot. Specifically, when the task is started, the Plot object creates the graphical element (eg. a Point) and then calls its update method with any data that is received through its Node.

Data-to-graphical mappings are defined in a data subdictionary, and additional parameters can be passed to the plot – in the below example, for example, a chance_bar is drawn as a horizontal line across the plot. By default it is drawn at 0.5, but its height can be set with an additional parameter chance_level. Available graphical primitives are registered in the plots.PLOT_LIST, and additional parameters are documented in the Plot class.

Data is plotted either by trial (default) or by timestamp (if PLOT['continuous'] != True). Numerical data is plotted (on the y-axis) as expected, but further mappings can be defined by extending the graphical element’s update method – eg. ‘L’(eft) maps to 0 and ‘R’(ight) maps to 1 by default.

PLOT = {
    'data': {
        'target'   : 'point',
        'response' : 'segment',
        'correct'  : 'rollmean'
    'chance_bar'  : True, # Draw a red bar at 50%
    'roll_window' : 50    # n trials to take rolling mean over

The above PLOT dictionary produces this pretty little plot:


screenshot of default nafc plot


The HARDWARE dictionary maps a hardware type (eg. POKES) and identifier (eg. 'L') to a Hardware object. The task uses the hardware parameterization in the prefs file (also see setup_pilot) to instantiate each of the hardware objects, so their naming system must match (ie. there must be a prefs.PINS['POKES']['L'] entry in prefs for a task that has a task.HARDWARE['POKES']['L'] object).

from autopilot.core import hardware

        'L': hardware.Beambreak,
        'C': hardware.Beambreak,
        'R': hardware.Beambreak
        'L': hardware.LED_RGB,
        'C': hardware.LED_RGB,
        'R': hardware.LED_RGB
        'L': hardware.Solenoid,
        'C': hardware.Solenoid,
        'R': hardware.Solenoid


First, the parameters that are given to the task when it is initialized are stored as attributes, either by unpacking **kwargs

class Nafc(Task):

    def __init__(**kwargs):
        for key, value in kwargs.items():
            setattr(self, key, value)

Or explicitly, which is recommended as it is more transparent:

class Nafc(Task):

    def __init__(self, stage_block=None, stim=None, reward=50, req_reward=False,
                 punish_stim=False, punish_dur=100, correction=False, correction_pct=50.,
                 bias_mode=False, bias_threshold=20, current_trial=0, **kwargs):

        self.req_reward     = bool(req_reward)
        self.punish_stim    = bool(punish_stim)
        self.punish_dur     = float(punish_dur)
        self.correction     = bool(correction)
        self.correction_pct = float(correction_pct)/100
        self.bias_mode      = bias_mode
        self.bias_threshold = float(bias_threshold)/100

        # etc...

Then the hardware is instantiated using a method inherited from the Task class:


Stimulus managers need to be instantiated separately. Currently, stimulus management details like correction trial percentage or bias correction are given as separate parameters, but will be included in the stim parameter in the future:

# use the init_manager wrapper to choose the correct stimulus manager
self.stim_manager = init_manager(stim)

# give the sounds a function to call when they end

if self.correction:

if self.bias_mode:

There are a few attributes that can be set at initialization that are unique:

  • stage_block - if the task is structured such that the Pilot calls each stage method and returns the resulting data, this threading.Event is used to wait between stages – an example will be shown below.

  • stages - an iterator or generator that yields stage methods.

In this example we have structured the task such that its stages (described below) are called in an endless cycle:

# This allows us to cycle through the task by just repeatedly calling
stage_list = [self.request, self.discrim, self.reinforcement]
self.stages = itertools.cycle(stage_list)

Stage Methods

The logic of a task is implemented in one or several stages. This example Nafc class uses three:

  1. request - precomputes the target and distractor ports, caches the stimulus, and sets the stimulus to play when the center port is entered

  2. discrim - sets the reward and punishment triggers for the target and distractor ports

  3. reinforcement - computes the trial result and readies the task for the next trial.

This task does not call its own stage methods, as we will see in the Wheel task example, but allows the Pilot to control them, and advances through stages using a stage_block that allows passage whenever a GPIO trigger is activated. Data is returned from each of the stage methods and is then returned to the Terminal by the Pilot.


First, the stage_block is cleared so that the task will not advance until one of the triggers is called. The target and distractor ports are yielded by the stim_manager along with the stimulus object.

def request(self, *args, **kwargs):
    # Set the event block

    # get next stim, self.distractor, self.stim = self.stim_manager.next_stim()
    # buffer it

Then triggers are stored under the name of the trigger (eg. ‘C’ for a trigger that comes from the center poke). All triggers need to be callable, and can be set either individually or as a series, as in this example. A lambda function is used to set a trigger with arguments – the center LED is set from green to blue when the stimulus starts playing.

A single task class can support multiple operating modes depending on its parameters. If the task has been asked to give request rewards (see Training a Subject), it adds an additional trigger to open the center solenoid.

# set the center light to green before the stimulus is played.
self.set_leds({'C': [0, 255, 0]})

# Set sound trigger and LEDs
# We make two triggers to play the sound and change the light color
change_to_blue = lambda: self.pins['LEDS']['C'].set_color([0,0,255])

# set triggers
if self.req_reward is True:
    self.triggers['C'] = [,
    self.triggers['C'] = [,

Finally, the data for this stage of the trial is gathered and returned to the Pilot. Since stimuli have variable numbers and names of parameters, both the table set up by the Subject and the data returning routine here extract stimulus parameters programmatically.

self.current_trial =
data = {
    'target'     :,
    'trial_num'  : self.current_trial,
    'correction' : self.correction_trial
# get stim info and add to data dict
sound_info = {k:getattr(self.stim, k) for k in self.stim.PARAMS}

return data

At the end of this function, the center LED is green, and if the subject pokes the center port the stimulus will play and then the next stage method will be called.

The center LED also turns from green to blue when the stimulus begins to play and then turns off when it is finished. This relies on additional methods that will be explained below.


The discrim method simply sets the next round of triggers and returns the request timestamp from the current trial. If either the target or distractor ports are triggered, the appropriate solenoid is opened or the punish method is called.

The trial_num is returned each stage for an additional layer of redundancy in data alignment.

def discrim(self,*args,**kwargs):
    # clear stage block to wait for triggers

    # set triggers
    self.triggers[]     = [lambda: self.respond(,
    self.triggers[self.distractor] = [lambda: self.respond(self.distractor),

    # Only data is the timestamp
    data = {'RQ_timestamp' :,
            'trial_num'    : self.current_trial}
    return data


pigpio can give us 5 microsecond measurement precision for triggers, currently we just use for timestamps, but highly accurate timestamps can be had by stashing the ticks argument given by pigpio to the handle_trigger() method. We will implement this if you don’t first :)


This method computes the results of the tasks and returns them with another timestamp. This stage doesn’t clear the stage_block because we want the next trial to be started immediately after this stage completes.

The results of the current trial are given to the stimulus manager’s update() method so that it can keep track of trial history and do things like bias correction, etc.

The TRIAL_END flag in the data signals to the Subject class that the trial is finished and its row of data should be written to disk. This, along with providing the trial_num on each stage, ensure that data is not misaligned between trials.

def reinforcement(self,*args,**kwargs):

    if self.response ==
        self.correct = 1
        self.correct = 0

    # update stim manager
    self.stim_manager.update(self.response, self.correct)

    data = {
        'DC_timestamp' :,
        'response'     : self.response,
        'correct'      : self.correct,
        'trial_num'    : self.current_trial,
        'TRIAL_END'    : True
    return data

Additional Methods

Autopilot doesn’t confine the logic of a task to its stage methods, instead users can use additional methods to give their task additional functionality.

These can range from trivial methods that just store values, such as the respond and stim_start methods:

def respond(self, pin):
    self.response = pin

def stim_start(self):
    self.discrim_playing = True

To more complex methods that operate effectively like stages, like the punish method, which flashes the LEDs and plays a punishment stimulus like white noise if it has been configured to do so:

def punish(self):
    # clear the punish block to the task doesn't advance while
    # punishment is delivered

    # if there is some punishment stimulus, play it
    if self.punish_stim:

    # flash LEDs and then clear the block once they are finished.
    threading.Timer(self.punish_dur / 1000.,

Additionally, since we gave the stimulus manager a trigger method that is called when the stimulus ends, we can turn the light blue when a stimulus is playing, and turn it off when it finishes

def stim_end(self):
    called by stimulus callback

    set outside lights blue
    # Called by the discrim sound's table trigger when playback is finished
    # Used in punishing leaving early
    self.discrim_playing = False
    #if not self.bailed and self.current_stage == 1:
    self.set_leds({'L':[0,255,0], 'R':[0,255,0]})

Distributed Go/No-Go - Using Child Agents


To demonstrate the use of Child agents, we’ll build the distributed Go/No-Go task described in section 4.3 of the Autopilot whitepaper.

In short, a subject runs on a circular running wheel whose velocity is measured by a laser computer mouse. When the subject ‘fixates’ by slowing below a threshold velocity, an drifting Gabor grating is presented. If the grating changes angles, the subject is rewarded if they lick in an IR beambreak sensor. If the grating doesn’t change angles, the subject is rewarded if they refrain from licking until the stimulus has ended.

Additional Prefs

To use a Child with this task, we will need to have a second Raspberry Pi setup with the same routine as a Pilot, except it needs the following values in its prefs.json file:

Child Prefs

    "NAME" : "wheel_child",
    "LINEAGE" : "CHILD",
    "PARENTID" : "parent_pilot",
    "PARENTIP" : "ip.of.parent.pilot",
    "PARENTPORT": "<MSGPORT of parent>",

And the parent pilot needs to have

Parent Prefs

    "NAME": "parent_pilot",
    "CHILDID": "wheel_child",

Go/No-Go Parameterization

The parameterization for this task is similar to that of the Nafc task above with a few extensions…

from autopilot.tasks import Task

class GoNoGo(Task):

    # Task parameterization
    PARAMS = odict()
    PARAMS['reward'] = {'tag': 'Reward Duration (ms)',
                        'type': 'int'}
    PARAMS['timeout']        = {'tag':'Delay Timeout (ms)',
    PARAMS['stim'] = {'tag':  'Visuals',
                      'type': 'visuals'}

    # Plot parameterization
    PLOT = {
        'data': {
            'x': 'shaded',
            'target': 'point',
            'response': 'segment'
        # our plot will use time as its x-axis rather than the trial number
        'continuous': True

    # TrialData descriptor
    class TrialData(tables.IsDescription):
        trial_num    = tables.Int32Col()
        target       = tables.BoolCol()
        response     = tables.StringCol(1)
        correct      = tables.Int32Col()
        RQ_timestamp = tables.StringCol(26)
        DC_timestamp = tables.StringCol(26)
        shift        = tables.Float32Col()
        angle        = tables.Float32Col()
        delay        = tables.Float32Col()

We add one additional data descriptor that describes the continuous data that will be sent from the Wheel object:

class ContinuousData(tables.IsDescription):
    x = tables.Float64Col()
    y = tables.Float64Col()
    t = tables.Float64Col()

The hardware specification is also similar, with one additional Flag object which behaves identically to the Beambreak object with reversed logic (triggered by 0->1 rather than 1->0).

    'POKES': {
        'C': hardware.Beambreak,
    'LEDS': {
        'C': hardware.LED_RGB,
    'PORTS': {
        'C': hardware.Solenoid,
    'FLAGS': {
        'F': hardware.Flag

Finally, we add an additional CHILDREN dictionary to specify the type of Child that we need to run the task, as well as any additional parameters needed to configure it.

The task_type must refer to some key in the autopilot.tasks.CHILDREN_LIST.


The Child agent is a subconfiguration of the Pilot agent, they will be delineated more explicitly as the agent framework is solidified.

    'WHEEL': {
        'task_type': "Wheel Child",


When initializing this task, we need to make our own Net_Node object as well as initialize our child. Assuming that the child is connected to the parent and appropriately configured (see the additional params above), then things should go smoothly.


Some of the parameters – most egregiously the Grating stimulus – are hardcoded in the initialization routine. This is bad practice but an unfortunately necessary evil because the visual stimulus infrastructure is not well developed yet.

from autopilot.stim.visual.visuals import Grating

def __init__(self, stim=None, reward = 50., timeout = 1000., stage_block = None,
             punish_dur = 500., **kwargs):
    super(GoNoGo, self).__init__()

    # we receive a stage_block from the pilot as usual, we won't use it
    # for task operation though.
    self.stage_block = stage_block
    self.trial_counter = itertools.count()

    # save parameters passed to us as arguments
    self.punish_dur = punish_dur
    self.reward = reward
    self.timeout = timeout
    self.subject = kwargs['subject']

    # init hardware and set reward as before

    # hardcoding stimulus while visual stim still immature
    self.stim = Grating(angle=0, freq=(4,0), rate=1, size=(1,1), debug=True)

    self.stages = itertools.cycle([self.request, self.discrim, self.reinforce])

Initializing the Net Node.

The Net_Node gets the following arguments:

  • id: The name that is used to identify the task’s networking object so other networking objects can send it messages. We prefix the pilot’s prefs.NAME with T_ because it is a task, though this is not required.

  • upstream: The name of the network node that is directly upstream from us, we will be sending our messages to the Pilot that is running us – and thus address it by its name

  • port: The port of our upstream mode, most commonly the prefs.MSGPORT

  • listens: A dictionary that maps messages with different ``KEY``s to specific handling methods. Since we don’t need to receive any data for this task, this is blank,

  • instance: Optional, denotes whether this node shouldn’t be the only node that exists within the Agent – ie. it uses the same instance of the tornado IOLoop as other nodes.

self.node = Net_Node(id="T_{}".format(prefs.NAME),

And then to initialize our Child we construct a message to send along to it.

Note that we send the message to prefs.NAME – we don’t want to have to know the IP address/etc. for our child because it connects to us – so the Station object handles sending it along with its Pilot_Station.l_child() listen.

# construct a message to send to the child
value = {
    'child': {'parent': prefs.NAME, 'subject': self.subject},
    'task_type': self.CHILDREN['WHEEL']['task_type'],
    'subject': self.subject

# send to the station object with a 'CHILD' key
self.node.send(to=prefs.NAME, key='CHILD', value=value)

The Child Task

The Wheel_Child task is a very thin wrapper around a Wheel object, which does most of the work.

It creates a stages iterator with a function that returns nothing to fit in with the general task structure.

class Wheel_Child(object):
    STAGE_NAMES = ['collect']

    PARAMS = odict()
    PARAMS['fs'] = {'tag': 'Velocity Reporting Rate (Hz)',
                    'type': 'int'}
    PARAMS['thresh'] = {'tag': 'Distance Threshold',
                        'type': 'int'}

    HARDWARE = {
        "OUTPUT": Digital_Out,
        "WHEEL":  Wheel

    def __init__(self, stage_block=None, fs=10, thresh=100, **kwargs):
        self.fs = fs
        self.thresh = thresh

        self.hardware = {}
        self.hardware['OUTPUT'] = Digital_Out(prefs.PINS['OUTPUT'])
        self.hardware['WHEEL'] = Wheel(digi_out = self.hardware['OUTPUT'],
                                       fs       = self.fs,
                                       thresh   = self.thresh,
                                       mode     = "steady")
        self.stages = cycle([self.noop])
        self.stage_block = stage_block

    def noop(self):
        # just fitting in with the task structure.
        return {}

    def end(self):

A Very Smart Wheel

Most of the Child’s contribution to the task is performed by the Wheel object.

The Wheel accesses a USB mouse connected to the Pilot, continuously collects its movements, and reports them back to the Terminal with a specified frequency (fs) with an internal Net_Node

An abbreviated version…

from inputs import devices

class Wheel(Hardware):

    def __init__(self, mouse_idx=0, fs=10, thresh=100, thresh_type='dist', start=True,
                 digi_out = False, mode='vel_total', integrate_dur=5):

        self.mouse = devices.mice[mouse_idx]
        self.fs = fs
        self.thresh = thresh
        # time between updates
        self.update_dur = 1./float(self.fs)

The Wheel has three message types,

  • 'MEASURE' - the main task is telling us to monitor for a threshold crossing, ie. previous trial is over and it’s ready for another one.

  • 'CLEAR' - stop measuring for a threshold crossing event!

  • 'STOP' - the task is over, clear resources and shut down.

    # initialize networking
    self.listens = {'MEASURE': self.l_measure,
                    'CLEAR'  : self.l_clear,
                    'STOP'   : self.l_stop}

    self.node = Net_Node('wheel_{}'.format(mouse_idx),

    # if we are being used in a child object,
    # we send our trigger via a GPIO pin
    self.digi_out = digi_out

    self.thread = None

    if start:

def start(self):
    self.thread = threading.Thread(target=self._record)
    self.thread.daemon = True

The wheel starts two threads, one that captures mouse movement events and puts them in a queue, and another that processes movements, transmits them to the Terminal, and handles the threshold triggers when the subject falls below a certain velocity.

def _mouse(self):
    # read mouse movements and put them in a queue
    while self.quit_evt:
        events =

def _record(self):


    # a threading.Event is used to terminate the wheel's operation
    while not self.quit_evt.is_set():

    # ... mouse movements are collected into a 2d numpy array ...

    # if the main task has told us to measure for a velocity threshold
    # we check if our recent movements (move) trigger the threshold
    if self.measure_evt.is_set():
        do_trigger = self.check_thresh(move)
        if do_trigger:

    # and we report recent movements back to the Terminal
    # the recent velocities and timestamp have been calculated as
    # x_vel, y_vel, and nowtime

If the threshold is triggered, a method (…``thresh_trig``…) is called that sends a voltage pulse through the Digital_Out given to it by the Child task.

def thresh_trig(self):
    if self.digi_out:

Go/No-Go Stage Methods

After the child is initialized, the Parent pilot begins to call the three stage functions for the task in a cycle

Very similar to the Nafc task above…

  • request - Tell the Child to begin measuring for a velocity threshold crossing, prepare the stimulus for delivery

  • discrim - Present the stimulus

  • reinforce - Reward the subject if they were correct

The code here has been abbreviated for the purpose of the example:

def request(self):
    # Set the event lock
    # wait on any ongoing punishment stimulus

    # set triggers
    self.triggers['F'] = [
        lambda:'shift', self.shift )

    # tell our wheel to start measuring
    self.node.send(to=[prefs.NAME, prefs.CHILDID, 'wheel_0'],

    # return data from current stage
    self.current_trial =
    data = {
        'target':, # whether to 'go' or 'not go'
        'shift': self.shift,   # how much to shift the
                               # angle of the stimulus
        'trial_num': self.current_trial

    return data

def discrim(self):
    # if the subject licks on a good trial, reward.
    # set a trigger to respond false if delay time elapses
        self.triggers['C'] = [lambda: self.respond(True), self.pins['PORTS']['C'].open]
        self.triggers['T'] = [lambda: self.respond(False), self.punish]
    # otherwise punish
        self.triggers['C'] = [lambda: self.respond(True), self.punish]
        self.triggers['T'] = [lambda: self.respond(False), self.pins['PORTS']['C'].open]

    # the stimulus has just started playing, wait a bit and then shift it (if we're gonna
    # choose a random delay
    delay = 0.0
    if self.shift != 0:
        delay = (random()*3000.0)+1000.0
        # a delay timer is set that shifts the stimulus after
        # <delay> milliseconds
        self.delayed_set(delay, 'shift', self.shift)

    # trigger the timeout in 5 seconds
    self.timer = threading.Timer(5.0, self.handle_trigger, args=('T', True, None)).start()

    # return data to the pilot
    data = {
        'delay': delay,
        'trial_num': self.current_trial

    return data

def reinforce(self):

    # stop timer if it's still going
    except AttributeError:
    self.timer = None

    data = {
        'response': self.response,
        'correct': self.correct,
        'trial_num': self.current_trial,
        'TRIAL_END': True
    return data