Autopilot is an integrated system for coordinating all parts of an experiment, but it is also designed to be permissive about how it is used and to make transitioning from existing lab tooling gentler – so its modules can be used independently.

To get a sample of autopilot, you can check out some of its modules without doing a fully configured Installation . As you get more comfortable using Autopilot, adopting more of its modules and usage patterns makes integrating each of the separate modules simpler and more powerful, but we’ll get there in time.

Minimal Installation

Say you have a Raspberry Pi with Raspbian installed . Install autopilot and its basic system dependencies & configuration like this:

pip3 install auto-pi-lot
python3 -m autopilot.setup.run_script env_pilot pigpiod

Capture Video

Say you have a Raspberry Pi Camera Module , capture some video! First make sure the camera is enabled:

python3 -m autopilot.setup.run_script picamera

and then capture a video with cameras.PiCamera and write it to test_video.mp4:

from autopilot.hardware.cameras import PiCamera
cam = PiCamera(name="my_picamera", fps=30)


Since every hardware object in autopilot is by default nonblocking (eg. work happens in multiple threads, you can make other calls while the camera is capturing, etc.), this will work in an interactive python session but would require that you sleep or call cam.stoppping.join() or some other means of keeping the process open.

While the camera is capturing, you can access its current frame in its frame attribute, or to make sure you get every frame, by calling queue() .

Communicate Between Computers

Synchronization and coordination of code across multiple computers is a very general problem, and an increasingly common one for neuroscientists as we try to combine many hardware components to do complex experiments.

Say our first raspi has an IP address and we get another raspi whose IP is . We can send messages between the two using two networking.Net_Node s. networking.Net_Node s send messages with a key and value , such that the key is used to determine which of its listens methods/functions it should call to handle value .

For this example, how about we make pilot 1 ping pilot 2 and have it respond with the current time?

On pilot 2, we make a node that listens for messages on port 5000. The upstream and port arguments here don’t matter since this node doesn’t initiate any connection, just received them (we’ll use a global variable here and hardcode the return id since we’re in scripting mode, but there are better ways to do this in autopilot proper):

from autopilot.networking import Net_Node
from datetime import datetime
global node_2

def thetime(value):
    global node_2
        to='pilot_1', key='THETIME',

node_2 = Net_Node(
    id='pilot_2', router_port=5000, upstream='', port=9999,

On pilot 1, we can then make a node that connects to pilot 2 and prints the time when it receives a response:

from autopilot.networking import Net_Node

node_1 = Net_Node(
    id='pilot_1', upstream='pilot_2',
    port=5000, upstream_ip = '',
    listens = {'THETIME':print}

node_1.send(to='pilot_1', key='WHATIS')

Realtime DeepLabCut

Autopilot integrates DeepLabCut-Live [KLS+20] ! You can use your own pretrained models (stored in your autopilot user directory under /dlc ) or models from the Model Zoo .

Now let’s say we have a desktop linux machine with DeepLabCut and dlc-live installed. DeepLabCut-Live is implemented in Autopilot with the transform.image.DLC object, part of the transform module.

First, assuming we have some image img (as a numpy array), we can process the image to get an array of x,y positions for each of the tracked points:

from autopilot import transform as t
import numpy as np

dlc = t.image.DLC(model_zoo='full_human')
points = dlc.process(img)

Autopilot’s transform module lets us compose multiple data transformations together with + to make deploying chains of computation to other computers. How about we process an image and determine whether the left hand in the image is raised above the head?:

# select the two body parts, which will return a 2x2 array
dlc += t.selection.DLCSlice(select=('wrist1', 'forehead'))

# slice out the 1st column (y) with a tuple of slice objects
dlc += t.selection.Slice(select=(

# compare the first (wrist) y position to the second (forehead)
dlc += t.logical.Compare(np.greater)

# use it!

Put it Together - Close a Loop!

We’ve tried a few things, why not put them together?

Let’s use our two raspberry pis and our desktop GPU-bearing computer to record a video of someone and turn an LED on when their hand is over their head. We could do this two (or one) computer as well, but let’s be extravagant.

Let’s say pilot 1, pilot 2, and the gpu computer have ip addresses of,, and, respectively.

Pilot 1 - Image Capture

On pilot 1, we configure our PiCamera to stream to the gpu computer. While we’re at it, we might as well also save a local copy of the video to watch later. The camera won’t stop capturing, streaming, or writing until we call capture():

from autopilot.hardware.cameras import PiCamera
cam = PiCamera()
cam.stream(to='gpu', ip='', port=5000)

GPU Computer

On the gpu computer, we need to receive frames, process them with the above defined transformation chain, and send the results on to pilot 2, which will control the LED. We could do this with the objects that we’ve already seen (make the transform object, make some callback function that sends a frame through it and give it to a Net_Node as a listen method), but we’ll make use of the Transformer “child” object – which is a peculiar type of Task designed to perform some auxiliary function in an experiment.

Rather than giving it an already-instantiated transform object, we instead give it a schematic representation of the transform to be constructed – When used with the rest of autopilot, this is to both enable it to be dispatched flexibly to different computers, but also to preserve a clear chain of data provenance by keeping logs of every parameter used to perform an experiment.

The Transformer class uses make_transform() to reconstitute it, receives messages containing data to process, and then forwards them on to some other node. We use its trigger mode, which only sends the value on to the final recipient with the key 'TRIGGER' when it changes.:

from autopilot.tasks.children import Transformer
import numpy as np

transform_description = [
        "transform": "DLC",
        "kwargs": {'model_zoo':'full_human'}
        "transform": "DLCSlice",
        "kwargs": {"select": ("wrist1", "forehead")}
        "transform": "Slice",
        "kwargs": {"select":(
        "transform": "Compare",
        "args": [np.greater],

transformer = Transformer(
    transform = transform_description
    operation = "trigger",
    node_id = "gpu",
    return_id = 'pilot_2',
    return_ip = '',
    return_port = 5001,
    return_key = 'TRIGGER',
    router_port = 5000

Pilot 2 - LED

And finally on pilot 2 we just write a listen callback to handle the incoming trigger:

from autopilot.hardware.gpio import Digital_Out
from autopilot.networking.Net_Node

global led
led = Digital_Out(pin=7)

def led_trigger(value:bool):
    global led

node = Net_Node(
    id='pilot_2', router_port=5001, upstream='', port=9999,
    listens = {'TRIGGER':led_trigger}

There you have it! Just start capturing on pilot 1:


What Next?

The rest of Autopilot expands on this basic use by providing tools to do the rest of your experiment, and to make replicable science easy.

  • write standardized experimental protocols that consist of multiple Task s linked by flexible graduation criteria

  • extend the library to use your custom hardware, and make your work available to anyone with our plugins system integrated with the autopilot wiki

  • Use our GUI that makes managing many experimental rigs simple from a single computer.

and so on…